HomeMaterial review · 2026-10-21
What is ready to practice, and what still needs work?
Use the scoped examples now. Check the limits before choosing a larger project, and help us improve the material.
104 lesson CPU references and 28 project reference checkers match the current sources. 0 independent expert teaching reviews and 0 learner pilots are recorded. A reference pass does not establish comprehension or professional readiness.
Read the labels
Available draft: the scoped practice material exists. Teaching review and learner testing are still pending.
In development: useful material exists, with important work still missing from the course or project.
Planned: full capstone implementation and qualification are pending. A separately scoped CPU pilot may be available and is reported as its own evidence.
Needs repair: required material is missing or a current check fails.
These describe the material. They do not change your saved progress, quiz results or portfolio evidence. Lesson and project runtimes have their own boundaries. Study-time estimates are not calibrated learner completion times.
Useful contributions now
Start with one changed example, unclear explanation, annotated figure or reproducible bug. Use the contribution link beside an item to open a GitHub issue with its context. For a code or writing change, propose a focused pull request against its canonical source.
- Run a consented novice walkthrough and record where the instructions need help.
- Review mathematical assumptions, numerical checks, figure interpretations and task rubrics.
- Add feedback for specific wrong checkpoint answers and changed-condition learner tests.
- Build one realistic model lineage with licensed data, frozen evaluation and exported reload.
- Review career coverage as a practitioner and explain what the portfolio does not establish.
See the teaching review protocol and complete audit with every exercise and project stage. This internal audit is not an independent human review.
What the hardware evidence establishes
4 source-matching physical v5e target receipts are recorded. They qualify named experiments, not an entire course, production model or TPU generation.
- v5e1-training-recovery-precision-20261020: One v5e chip: bounded training, resume and single-device dot/precision checks; no multi-host or model-quality qualification.
- v5e1-pallas-20261020: One v5e chip, two kernels, FP32/BF16, logical shape257x513 and five timing samples; not a general speed or scaling claim.
- v5e4-collectives-20261020: One v5e host with four devices: deterministic psum, reduce-scatter and all-gather numerical, placement, StableHLO and trace checks; not multi-host or scaling qualification.
- v5e16-multihost-checkpoint-20261020: Four VM controllers and sixteen v5e devices: process-local synthetic data, global updates, Orbax GCS save at step4, same-process abstract sharded restore and replay through step8. No worker failure, fresh-process restart, elastic recovery, production quality or scaling measurement.
Fresh-process or worker-loss multi-host recovery, general scaling, production-model quality and edge-device deployment need their own evidence.
Course phases
Setup & first steps
Available draftA concrete saved-file and environment walkthrough, plus CPU/device checks.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Run consented fresh-machine novice pilots on Windows, macOS and Linux; record where learners need help.
Arrays & pure functions
Available draftDense-array axes, preprocessing and update contracts have changed-input checks.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Add a masked variable-length batch that continues into the model project; review constant-feature and ragged-input explanations.
JAX transformations
Available draftGradients, batching, compilation and tracing have independent arithmetic and changed conditions.
Evidence: 5/5 current CPU references. Teaching review pending.
Help needed: Review transformation composition at nonsmooth boundaries and add a single continuing model across the five lessons.
State, randomness & control flow
Available draftKey ownership, parameter trees, scan carries and branches have explicit invariants.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Connect sample IDs, augmentation keys and optimizer state to a recovery handoff; test that novices can explain key reuse.
Math & optimization
Available draftTwelve connected math labs and an optimizer audit cover geometry, conditioning, noise and update state.
Evidence: 12/12 current CPU references. Teaching review pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Neural network training
Available draftTiny MLP, NNX and CNN references now include real multilayer failure localization, named state identity and BatchNorm/Dropout evaluation isolation.
Evidence: 5/5 current CPU references. Teaching review pending.
Help needed: Add a licensed natural-image pilot with frozen source-group splits, a simple baseline and per-class error slices.
Data & checkpoint recovery
Available draftGrain/Orbax and the recovery audit exercise complete state and input-position contracts.
Evidence: 6/6 current CPU references. Teaching review pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Transformers & language models
In developmentAttention and packing lessons connect to a separate configurable multi-layer decoder with train-only BPE, a measured natural-text CPU pilot, cache parity and fresh-process full-state recovery.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Review tokenizer and length policies on broader licensed corpora, extend the measured short pilot into staged training and evaluate generation quality. Target-scale training and independent teaching review remain open.
Performance diagnosis
Available draftWarm timing, call ledgers, retained CPU traces and independent memory estimates are available.
Evidence: 5/5 current CPU references. Teaching review pending.
Help needed: Annotate a real target trace with the exact workload and bottleneck; retain distributions and distinguish estimates from measured allocated memory.
Distributed training
In developmentLogical CPU references include an oracle-checked 2x2 parameter/momentum-sharded update with unequal and empty valid counts; physical target evidence remains separately scoped.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Qualify fresh-process multi-host restart and worker-loss recovery with explicit data ownership; do not infer scaling results from logical CPU devices.
Scientific computing
In developmentFixed-grid ODE references connect to a separately pinned adaptive Diffrax extension with stiff solves, tolerance-sensitive gradients, noisy inversion and identifiable versus insensitive profiles.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Review the solver and inference assumptions with a domain specialist, extend the synthetic fixture to real measurements and validate uncertainty or model misspecification before broader scientific claims.
Probabilistic modeling
In developmentExact Bayesian regression, instructional HMC/VI and an optional real NumPyro workflow with raw chains, modern diagnostics and paired model-mismatch checks.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Review the probabilistic assumptions with a specialist and extend the synthetic model checks to a licensed real-data case with calibrated predictive evaluation.
Reinforcement learning
In developmentTabular policy objectives, paired-seed evaluation, a separately learned value table and explicit GAE terminal/cutoff/reset checks have bounded CPU references.
Evidence: 5/5 current CPU references. Teaching review pending.
Help needed: Connect the learned critic and GAE to actor updates, then evaluate larger tasks or continuous actions with independent learner and expert review.
Pallas kernels
In developmentCPU interpretation and a separately qualified two-kernel v5e target are available.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Turn the bounded target receipt into a learner tuning lab with retained trace interpretation and end-to-end comparison; GPU and general pipeline claims remain unqualified.
Autodiff & JAX internals
Available draftJaxpr, directional products, custom rules and a supported-subset interpreter have independent derivative checks.
Evidence: 4/4 current CPU references. Teaching review pending.
Help needed: Add transformation-composition failures and effect/extension boundaries; have an independent reviewer check the supported subset and numerical assumptions.
Deployment, interoperability & edge AI
In developmentDense and small saved-decoder PyTorch-to-NNX parity, shifted/masked sequence semantics, fresh-process artifacts and quantization contracts have bounded CPU checks.
Evidence: 8/8 current CPU references. Teaching review pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Workload operations
In developmentLocal job incidents, reloaded-prediction slice gates, executed local retrieval traces, behavioral rollback and identity-bound retirement drills are runnable teaching fixtures.
Evidence: 8/8 current CPU references. Teaching review pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
Self-supervised pretraining: masked and contrastive learning
In developmentMasked-token, masked-image and contrastive mechanics now connect to saved small encoders, held-out masks/groups and frozen probes with explicit failed-baseline limitations.
Evidence: 3/3 current CPU references. Teaching review pending.
Help needed: Extend saved-encoder teaching fixtures to licensed data recipes and larger architectures with independent representation evaluation; the full modality capstones remain planned.
Post-training: SFT, LoRA, reward models and RLHF
In developmentResponse masks, base-bound LoRA artifacts, held-out synthetic reward checks and bounded saved-decoder SFT/DPO handoffs accompany the separately recorded failed-quality PPO extension.
Evidence: 5/5 current CPU references. Teaching review pending.
Help needed: The saved-decoder extension failed its strict heldout reward quality floor. Human preference quality, retained behavior and target-quality evaluation remain open and need separately reviewed evidence.
TPU Setup: Provisioning & Runtime Verification
In developmentProvisioning guidance and runtime verification connect to separate bounded physical v5e evidence.
Evidence: 2/2 current CPU references. Teaching review pending.
Help needed: Review generation-specific topology and memory assumptions, test a learner provisioning journey and keep pricing illustrative rather than a live quote.
TPU Workflows: Launching Jobs & Checkpointed Experiments
In developmentReviewed launch targets, checkpoint contracts and a bounded multi-host receipt exist.
Evidence: 2/2 current CPU references. Teaching review pending.
Help needed: Qualify fresh-process and worker-loss recovery, retained per-worker evidence and cleanup in an observed learner run; device discovery alone is not training qualification.
TPU Systems: Generations, Memory, Precision & XProf
In developmentPer-device training/inference memory ledgers and policy-specific precision bytes connect to stable warmed profiling and an actual CPU trace-reading example; physical targets remain narrowly qualified.
Evidence: 2/2 current CPU references. Teaching review pending.
Help needed: Add retained target-specific trace interpretation and native low-bit kernel/peak-HBM measurements. CPU event accounting and dequantized references do not qualify those target claims.
Lessons and their exercises
Set up your learning workspace
Available draftDownload a starter workspace, select its Python environment, run a saved program, and verify one deliberate change.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Run consented fresh-machine novice pilots on Windows, macOS and Linux; record where learners need help.
Meet your arrays and devices
Available draftCreate an array, inspect its dtype and device, and explain the result.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Run consented fresh-machine novice pilots on Windows, macOS and Linux; record where learners need help.
Practice with four virtual CPU devices
Available draftStart four logical CPU devices, inspect row-sharded and replicated arrays, and diagnose an indivisible batch.
Evidence: Logical CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Run consented fresh-machine novice pilots on Windows, macOS and Linux; record where learners need help.
Move your experiment to a TPU
Available draftUse the recorded CPU reference to check arithmetic and placement, then run only a named, source-reviewed TPU target with pinned dependencies and verify its observed backend receipt.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Run consented fresh-machine novice pilots on Windows, macOS and Linux; record where learners need help.
From NumPy to jax.numpy
Available draftBuild feature-wise preprocessing with fitted statistics and verify behavior on new and constant data.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a masked variable-length batch that continues into the model project; review constant-feature and ragged-input explanations.
Shapes, broadcasting, and dtypes
Available draftSpecify broadcasting contracts and detect silent pairwise-loss and dtype mistakes.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a masked variable-length batch that continues into the model project; review constant-feature and ragged-input explanations.
Pure functions and explicit inputs
Available draftDesign an explicit parameter/data interface and return new state without mutating old containers.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a masked variable-length batch that continues into the model project; review constant-feature and ragged-input explanations.
Immutable updates and indexing
Available draftImplement functional indexed updates and distinguish accumulated contributions from shape-preserving masks.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a masked variable-length batch that continues into the model project; review constant-feature and ragged-input explanations.
Your first gradient
Available draftExplain local sensitivity and the chain rule, then compare analytic, numerical, and automatic derivatives.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review transformation composition at nonsmooth boundaries and add a single continuing model across the five lessons.
Losses and value_and_grad
Available draftDerive and audit a scalar loss, its parameter gradient, and stable versus unstable updates.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review transformation composition at nonsmooth boundaries and add a single continuing model across the five lessons.
Batch a function with vmap
Available draftSpecify batching axes and connect per-example gradients to a reduced batch objective.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review transformation composition at nonsmooth boundaries and add a single continuing model across the five lessons.
Compile a function with jit
Available draftValidate compiled loss/gradient values and produce a correctly synchronized latency report.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review transformation composition at nonsmooth boundaries and add a single continuing model across the five lessons.
Tracing, static arguments, and recompilation
Available draftExplain the staging boundary and repair configuration versus runtime-control-flow mistakes.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review transformation composition at nonsmooth boundaries and add a single continuing model across the five lessons.
Random keys without surprises
Available draftSplit keys and show reproducibility without accidental key reuse.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect sample IDs, augmentation keys and optimizer state to a recovery handoff; test that novices can explain key reuse.
Pytrees and structured parameters
Available draftRepresent parameters and state in a nested structure and transform its leaves.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect sample IDs, augmentation keys and optimizer state to a recovery handoff; test that novices can explain key reuse.
Compiled loops with lax.scan
Available draftReplace a fixed-length Python loop with scan and compare final states.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect sample IDs, augmentation keys and optimizer state to a recovery handoff; test that novices can explain key reuse.
Branches with lax.cond
Available draftImplement a data-dependent branch that works inside a compiled function.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect sample IDs, augmentation keys and optimizer state to a recovery handoff; test that novices can explain key reuse.
Vectors, norms, and projections
Available draftRead a dot product geometrically and separate a vector into a projection and an orthogonal residual.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Linear algebra and loss intuition
Available draftConnect a matrix-vector product to predictions and a scalar regression loss.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Least squares, rank, and conditioning
Available draftFit a linear model with a direct solver and diagnose nonunique or sensitive coefficients.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
The chain rule, Jacobians, and directions
Available draftCompute a small Jacobian by hand and connect forward sensitivities to reverse-mode gradients.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Gradient checking and numerical accuracy
Available draftCompare autodiff with finite differences over several step sizes.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Curvature, Hessians, and learning rates
Available draftUse curvature to explain learning-rate stability, slow directions and why a zero gradient need not be a minimum.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Write gradient descent yourself
Available draftFit a regression model using explicit parameter updates.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Probability, likelihood, and stable losses
Available draftDerive binary cross-entropy from a probability model and compute it safely from logits.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Regularization and honest validation
Available draftDerive a ridge solution, distinguish data loss from penalized loss, and select a penalty using held-out data.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Minibatches, expectation, and gradient noise
Available draftExplain an unbiased minibatch gradient and verify how sampling and reduction affect its variance.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Optimize with Optax
Available draftReplace the handwritten update with an optimizer and track its state.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Adam, clipping, and learning-rate schedules
Available draftExplain Adam’s moments, inspect clipping and schedule order, and compare optimizers with a controlled update budget.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Pilot the synthesis with learners; add constrained or nonsmooth optimization only around a new inspectable task, and improve feedback on wrong answers.
Build a tiny multilayer perceptron
Available draftImplement a small model with explicit parameters before adding a model library.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a licensed natural-image pilot with frozen source-group splits, a simple baseline and per-class error slices.
Model and state with Flax NNX
Available draftExpress a model and inspect the distinction between parameters and other state.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a licensed natural-image pilot with frozen source-group splits, a simple baseline and per-class error slices.
A compiled train and evaluation step
Available draftBuild train and eval steps and check that evaluation does not update weights.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a licensed natural-image pilot with frozen source-group splits, a simple baseline and per-class error slices.
Train a CNN on a small dataset
Available draftTrain a classifier with a fixed data split and report held-out metrics.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a licensed natural-image pilot with frozen source-group splits, a simple baseline and per-class error slices.
Debug unstable learning
Available draftLocalize a multilayer model failure in data, activations, gradients or update scale, then repair one cause and compare held-out loss.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add a licensed natural-image pilot with frozen source-group splits, a simple baseline and per-class error slices.
Design an input pipeline
Available draftBuild and audit a deterministic batched host pipeline.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Load, batch, and checkpoint order with Grain
Available draftBuild a Grain source/sampler/batch pipeline and verify its record order.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Save model, optimizer, and random state
Available draftCheckpoint complete training state with Orbax and replay the next update.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Recover data position and resume
Available draftPersist model/optimizer/key and Grain iterator position at one boundary.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Asynchronous checkpoints and failure boundaries
Available draftDistinguish requested, completed, verified and published checkpoint state. Continue a functional update while an asynchronous writer is alive.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Track experiments and lineage with MLflow
Available draftTrack, retrieve and compare two JAX experiments using explicit data identity, metric steps and downloaded model artifacts.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review exact-continuation wording against its fixed environment; inject storage and concurrent-writer failures separately from ordinary local restore.
Attention from small pieces
Available draftImplement scaled dot-product attention and test its output shape.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review tokenizer and length policies on broader licensed corpora, extend the measured short pilot into staged training and evaluate generation quality. Target-scale training and independent teaching review remain open.
Masking, tokenization, and sequence packing
Available draftConstruct a causal mask and prevent packed sequences from attending across boundaries.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review tokenizer and length policies on broader licensed corpora, extend the measured short pilot into staged training and evaluate generation quality. Target-scale training and independent teaching review remain open.
A Transformer block and mixed precision
Available draftBuild a residual attention block and compare numerical behavior across precision choices.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review tokenizer and length policies on broader licensed corpora, extend the measured short pilot into staged training and evaluate generation quality. Target-scale training and independent teaching review remain open.
Train, checkpoint, and generate
Available draftTrain a tiny causal model, restore it, and generate from a fixed seed.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review tokenizer and length policies on broader licensed corpora, extend the measured short pilot into staged training and evaluate generation quality. Target-scale training and independent teaching review remain open.
Benchmark asynchronous work correctly
Available draftWarm up a computation and synchronize results before reporting timing.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Annotate a real target trace with the exact workload and bottleneck; retain distributions and distinguish estimates from measured allocated memory.
Diagnose recompilation and host synchronization
Available draftFind a shape-changing call or host read and connect it to observed overhead.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Annotate a real target trace with the exact workload and bottleneck; retain distributions and distinguish estimates from measured allocated memory.
Read an XProf trace
Available draftCapture an actual warmed CPU trace with named steps. Identify nested intervals and avoid double-counting their durations.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Annotate a real target trace with the exact workload and bottleneck; retain distributions and distinguish estimates from measured allocated memory.
Reason with HLO and the roofline model
Available draftLocate matrix multiplication and activation in StableHLO. Derive the operation and minimum byte counts for a declared matrix contract.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Annotate a real target trace with the exact workload and bottleneck; retain distributions and distinguish estimates from measured allocated memory.
Memory, rematerialization, and optimization tradeoffs
Available draftExplain why reverse-mode differentiation needs forward intermediates. Compare ordinary and rematerialized gradients with an independent reverse recurrence.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Annotate a real target trace with the exact workload and bottleneck; retain distributions and distinguish estimates from measured allocated memory.
Arrays, meshes, and sharding
Available draftPartition an array over a device mesh and inspect its local shards.
Evidence: Logical CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Qualify fresh-process multi-host restart and worker-loss recovery with explicit data ownership; do not infer scaling results from logical CPU devices.
A sharded training step
Available draftCompute global count-weighted gradients and verify a 2x2 data/model mesh forward, backward and momentum update against an unsharded oracle.
Evidence: Logical CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Qualify fresh-process multi-host restart and worker-loss recovery with explicit data ownership; do not infer scaling results from logical CPU devices.
Communication-efficient algorithms
Available draftCompare all-reduce and reduce-scatter against the same independent global gradient. Explain global result shape versus per-device storage.
Evidence: Logical CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Qualify fresh-process multi-host restart and worker-loss recovery with explicit data ownership; do not infer scaling results from logical CPU devices.
Resilient distributed training
Available draftRestore a complete training state with explicit target sharding. Compare next sample IDs, optimizer state and losses across an epoch boundary.
Evidence: Logical CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Qualify fresh-process multi-host restart and worker-loss recovery with explicit data ownership; do not infer scaling results from logical CPU devices.
A functional physical simulation
Available draftImplement a pure scalar state transition with units and a fixed time grid. Separate a discrete implementation oracle from continuous approximation error.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the solver and inference assumptions with a domain specialist, extend the synthetic fixture to real measurements and validate uncertainty or model misspecification before broader scientific claims.
Vectorize trajectories and scan time
Available draftConstruct and verify a fourth-order fixed-step transition. Distinguish batch-major from time-major trajectory layouts.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the solver and inference assumptions with a domain specialist, extend the synthetic fixture to real measurements and validate uncertainty or model misspecification before broader scientific claims.
Differentiate through a solver
Available draftDerive an independent discrete solver sensitivity. Check an observation-loss gradient at a nonoptimal parameter.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the solver and inference assumptions with a domain specialist, extend the synthetic fixture to real measurements and validate uncertainty or model misspecification before broader scientific claims.
Fit parameters and scale the simulation
Available draftFit a constrained physical parameter through an executed simulator. Separate training fit, held-out prediction and solver refinement.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the solver and inference assumptions with a domain specialist, extend the synthetic fixture to real measurements and validate uncertainty or model misspecification before broader scientific claims.
Distributions and sampling
Available draftDistinguish samples, density, variance and standard error Check normal moments against independent analytic expectations
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the probabilistic assumptions with a specialist and extend the synthetic model checks to a licensed real-data case with calibrated predictive evaluation.
Build a Bayesian regression
Available draftState the prior, likelihood and observation-noise assumptions Derive posterior precision and check a closed-form special case
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the probabilistic assumptions with a specialist and extend the synthetic model checks to a licensed real-data case with calibrated predictive evaluation.
Hamiltonian Monte Carlo and sampler diagnostics
Available draftImplement leapfrog with the correct half-step boundaries Check reversibility and an independent analytic gradient
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the probabilistic assumptions with a specialist and extend the synthetic model checks to a licensed real-data case with calibrated predictive evaluation.
Variational inference and model checking
Available draftDerive the diagonal Gaussian variational optimum for a correlated target Separate optimization error from variational-family error
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review the probabilistic assumptions with a specialist and extend the synthetic model checks to a licensed real-data case with calibrated predictive evaluation.
Write a functional environment
Available draftImplement reset and step as pure functions with explicit state. Distinguish termination, time-limit truncation and an absorbing finished state.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect the learned critic and GAE to actor updates, then evaluate larger tasks or continuous actions with independent learner and expert review.
Batch environments and scan rollouts
Available draftBuild a scan over vmapped transitions with explicit key ownership. Reconstruct episode returns and reward-to-go from padded trajectories.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect the learned critic and GAE to actor updates, then evaluate larger tasks or continuous actions with independent learner and expert review.
Policy gradients and PPO
Available draftDerive and implement the score-function policy gradient for complete episodes. Train with fresh stochastic rollouts and a fixed behavior policy per PPO batch.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect the learned critic and GAE to actor updates, then evaluate larger tasks or continuous actions with independent learner and expert review.
Evaluate agents across seeds
Available draftSeparate training seeds, evaluation streams and individual episodes. Report per-agent means and variability with an explicit aggregation unit.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect the learned critic and GAE to actor updates, then evaluate larger tasks or continuous actions with independent learner and expert review.
Learn value functions and generalized advantage estimation
Available draftFit sampled discounted return targets at gamma=0.9 over at most eight actions, then estimate GAE at gamma=0.9 and lambda=0.95 with terminal and collector-cutoff semantics explicit.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Connect the learned critic and GAE to actor updates, then evaluate larger tasks or continuous actions with independent learner and expert review.
Pallas grids and BlockSpecs
Available draftImplement an actual Pallas Ref-writing kernel with grid and BlockSpecs. Derive block coordinates and distinguish them from element offsets.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Turn the bounded target receipt into a learner tuning lab with retained trace interpretation and end-to-end comparison; GPU and general pipeline claims remain unqualified.
A first TPU kernel
Available draftImplement a real TPU-targeted Pallas launch with a broadcast BlockSpec. Specify target layout and float32 compute/output-cast policies.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Turn the bounded target receipt into a learner tuning lab with retained trace interpretation and end-to-end comparison; GPU and general pipeline claims remain unqualified.
Tiling, memory, and pipelining
Available draftExecute actual emit_pipeline semantics under explicit CPU TPU-layout simulation. Compare synchronous and buffered schedules against an independent oracle.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Turn the bounded target receipt into a learner tuning lab with retained trace interpretation and end-to-end comparison; GPU and general pipeline claims remain unqualified.
Iterate with correctness and performance evidence
Available draftBuild a representative shape/dtype correctness matrix for an actual Pallas pipeline. Separate kernel implementation error from input/output precision differences.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Turn the bounded target receipt into a learner tuning lab with retained trace interpretation and end-to-end comparison; GPU and general pipeline claims remain unqualified.
Read your first jaxpr
Available draftInspect a small transformed function and identify its operations.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add transformation-composition failures and effect/extension boundaries; have an independent reviewer check the supported subset and numerical assumptions.
JVPs, VJPs, and higher-order derivatives
Available draftCompute and interpret JVP and VJP products with correct domains and shapes. Verify the adjoint dot-product identity against a hand-derived Jacobian.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add transformation-composition failures and effect/extension boundaries; have an independent reviewer check the supported subset and numerical assumptions.
Custom derivative rules
Available draftImplement actual custom_jvp and custom_vjp rules with correct return and residual contracts. Verify seed linearity, stable extreme values and second derivatives.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add transformation-composition failures and effect/extension boundaries; have an independent reviewer check the supported subset and numerical assumptions.
Lowering, compilation, and a tiny transformation
Available draftInterpret literals, captured constants and variables in a supported flat jaxpr. Implement actual forward derivative rules without calling jax.jvp inside the interpreter.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add transformation-composition failures and effect/extension boundaries; have an independent reviewer check the supported subset and numerical assumptions.
Adapt a pretrained model and choose a post-training objective
Available draftPretrain and reload an artifact with recorded data and checkpoint hashes Compare supervised adaptation and teacher distillation from the same starting weights
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Move models between JAX, Keras, TensorFlow and PyTorch
Available draftVerify parameter layouts and inference semantics before switching frameworks.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Convert PyTorch weights to Flax and locate numerical errors
Available draftMap actual PyTorch parameters to Flax NNX and localize the first layer or gradient mismatch, then transfer to the optional small saved decoder.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Export a computation and verify its serving contract
Available draftRound-trip an exported inference function and reject incompatible requests.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Choose weight, activation and accumulation precision
Available draftCompare floating-point policies, weight-only quantization and calibrated integer activations.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Containerize a model service and verify its boundary
Available draftPackage an inference service and verify model identity, request validation and readiness across process and container boundaries.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Inference capacity, batching, and autoscaling
Available draftMeasure completed inference after warming each batch shape Report batch latency and example throughput with correct units
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Deploy at the edge: conversion, budgets and device checks
Available draftDefine an edge runtime contract and measure an end-to-end batch-one request.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Execute and document a real container service, concurrent request workload and named edge-device/delegate report; dense parity is not full architecture portability.
Accelerator jobs and runtime lifecycle
Available draftLaunch and reap a real bounded subprocess with explicit runtime requirements. Distinguish successful completion, worker failure and supervisor timeout.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
Health, logs, and workload observability
Available draftBuild useful summaries from structured events emitted by a real worker. Define throughput with explicit numerator, denominator and synchronization boundary.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
Failure recovery and operational runbooks
Available draftRehearse a real worker failure and resume from an atomic complete-state checkpoint. Verify the next update and every later state against an uninterrupted reference.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
Changes, rollback, and artifact provenance
Available draftActivate and roll back identified local model bundles, verify consumer predictions and preserve the active model after a rejected selection.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
Capacity, utilization, and operating cost
Available draftMeasure end-to-end local service time at more than one workload length. Compute offered load, worker reserve and illustrative cost with consistent units.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
MLOps: data contracts, CI gates and monitoring
Available draftEvaluate predictions from a reloaded local artifact and apply group-safe, count-weighted slice and release rules.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
LLMOps: version prompts, evaluate behavior and trace requests
Available draftTrace an executed local retrieval, prompt and answer workflow, pair responses by case ID and diagnose component failures before comparing candidates.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
ModelOps: own, approve, roll back and retire a model
Available draftBind approval to a model bundle, verify consumer activation/rollback, and enforce retired-target and fallback policies.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add authenticated approvals, multi-writer release control, load-driven SLO incidents and application-specific LLM evaluation beyond deterministic fixtures.
Masked language modeling: predict hidden tokens
Available draftAudit selected-token MLM loss and transfer a saved small encoder to held-out masks and a frozen representation probe.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Extend saved-encoder teaching fixtures to licensed data recipes and larger architectures with independent representation evaluation; the full modality capstones remain planned.
Masked image modeling: reconstruct missing patches
Available draftAudit visible-only masked reconstruction and compare a saved small encoder across changed masks and a frozen representation probe.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Extend saved-encoder teaching fixtures to licensed data recipes and larger architectures with independent representation evaluation; the full modality capstones remain planned.
Contrastive learning: views, positives and negatives
Available draftTrain a bounded contrastive objective and check held-out grouped retrieval with multiple positives and a saved encoder identity.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Extend saved-encoder teaching fixtures to licensed data recipes and larger architectures with independent representation evaluation; the full modality capstones remain planned.
Supervised fine-tuning with response-only token loss
Available draftAlign response-only shifted targets, then apply a bounded saved-decoder SFT transfer with frozen baselines, tokenizer identity and held-out reporting.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: The saved-decoder extension failed its strict heldout reward quality floor. Human preference quality, retained behavior and target-quality evaluation remain open and need separately reviewed evidence.
LoRA: adapt, save and merge low-rank updates
Available draftSeparate frozen-base and low-rank updates, then save and reload a base-bound adapter with nonunit scaling and wrong-base rejection.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: The saved-decoder extension failed its strict heldout reward quality floor. Human preference quality, retained behavior and target-quality evaluation remain open and need separately reviewed evidence.
Learn a reward model from pairwise preferences
Available draftFit pairwise reward differences and evaluate unchanged held-out unanimous and disagreement cases against a pre-fit baseline.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: The saved-decoder extension failed its strict heldout reward quality floor. Human preference quality, retained behavior and target-quality evaluation remain open and need separately reviewed evidence.
RLHF mechanics: a frozen reward and PPO policy updates
Available draftKeep three identities separate. Collect actions before optimizing them.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: The saved-decoder extension failed its strict heldout reward quality floor. Human preference quality, retained behavior and target-quality evaluation remain open and need separately reviewed evidence.
Direct preference optimization and reference-corrected margins
Available draftCompare reference-corrected preference margins, then score unequal-length saved-decoder responses with explicit masks and a bounded sequence update.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: The saved-decoder extension failed its strict heldout reward quality floor. Human preference quality, retained behavior and target-quality evaluation remain open and need separately reviewed evidence.
Select a TPU slice and provision your Cloud TPU VM
Available draftChoose and identify a named single-host TPU topology, distinguish physical chips from JAX devices, compare per-device memory-fit assumptions and verify a requested-target receipt before a route-owned teardown.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review generation-specific topology and memory assumptions, test a learner provisioning journey and keep pricing illustrative rather than a live quote.
Bootstrap .venv and verify a TPU runtime receipt
Available draftUse the pinned TPU runtime, compare synchronized arithmetic with its oracle and distinguish requested backend/device readiness from a CPU rehearsal; retrieve the receipt and clean up only named owned resources.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Review generation-specific topology and memory assumptions, test a learner provisioning journey and keep pricing illustrative rather than a live quote.
Stage and launch checkpointed TPU jobs
Available draftStage launch.py to a TPU VM, separate first-step XLA compilation warmup from steady-state step timing, and verify that resuming from an atomic checkpoint matches an uninterrupted run.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Qualify fresh-process and worker-loss recovery, retained per-worker evidence and cleanup in an observed learner run; device discovery alone is not training qualification.
Review scripts before TPU staging
Available draftClassify the exact program before staging: keep
distributed-02on CPU and generate a TPU launch only for a reviewed target whose backend, dependencies, topology, and data ownership are explicit. Distinguish multi-host discovery from a qualified distributed update.Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Qualify fresh-process and worker-loss recovery, retained per-worker evidence and cleanup in an observed learner run; device discovery alone is not training qualification.
TPU generations and HBM memory budgeting
Available draftCompare TPU generation tradeoffs, prove per-device replicated versus sharded memory fit, and budget training state separately from inference KV cache.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add retained target-specific trace interpretation and native low-bit kernel/peak-HBM measurements. CPU event accounting and dequantized references do not qualify those target claims.
bfloat16, int8 precision, and XProf profiling on TPU
Available draftMeasure BF16 operand and requested result-dtype error against a finite-precision FP32 JAX reference, inspect CPU dtype and compiler evidence, and retain a warmed XProf trace with its observed backend.
Evidence: CPU reference checked. Expert review pending; learner pilot pending.
Help needed: Add retained target-specific trace interpretation and native low-bit kernel/peak-HBM measurements. CPU event accounting and dequantized references do not qualify those target claims.
Runnable practice projects
Build a regression training audit
Available draftIndependent regression gradients and changed-condition training.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Pilot the rubric with unseen data and retain a reviewer-scored evidence package.
Build a reproducible classifier audit
Available draftSmall synthetic classifier with objective, gradient and replay checks.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Bridge to a licensed real dataset with frozen splits, error analysis and reviewer feedback.
Audit a differentiable scientific inverse problem
Available draftFixed-grid simulation plus an optional executed Diffrax workflow with adaptive/stiff solves, sensitivity refinement, same-data noisy fitting and explained identifiability figures.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add licensed real measurements, an explicit uncertainty or discrepancy model and domain-expert review. Existing evidence is synthetic CPU computation.
Train and audit a policy from interaction
Available draftTabular environment, policy objectives, held-out evaluation and a separately checked learned value table with GAE.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Integrate the critic/GAE with actor updates and evaluate a larger task with a fixed held-out protocol and learner review.
Audit a Bayesian regression model
Available draftExact regression oracle and an optional, separately pinned NumPyro NUTS workflow with raw-chain diagnostics and paired predictive checks.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Validate assumptions, diagnostics and model mismatch interpretation with specialist and learner review, then extend to a real-data example.
Audit derivatives and compile a tiny transformation
Available draftDirectional products, custom rules and a small primitive transformation.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Review higher-order composition and unsupported-primitive failure cases.
Release a trained computation with a deployment audit
Available draftSmall artifact adaptation, precision and request capacity contracts.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Attach a real container run, load test and target runtime report.
Operate and recover a local JAX workload
Available draftReal local process incidents, complete recovery and rollback.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Extend from local drills to authenticated, concurrent cloud release and SLO failure cases.
Build an audio waveform-to-inference harness
Available draftSynthetic waveform classification with recovery, precision and export.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add licensed recordings, speaker/source-held-out splits and task-appropriate baselines; this is not an ASR model.
Build and audit image–text retrieval
Available draftSmall paired-branch contrastive system with recovery and export.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add natural paired data, duplicate-positive handling and held-out multimodal retrieval.
Audit a Pallas kernel and its target evidence
Available draftInterpreted semantics plus two separately observed v5e target kernels.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Reproduce the bounded target on new shapes and interpret the trace before claiming a useful speedup.
Build a recoverable image classifier through exported inference
Available draftSynthetic CNN lifecycle with augmentation, recovery and frozen errors.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add real photos, source-group leakage checks and an observed edge/runtime report.
Connect arrays, transformations and explicit state
Available draftEnvironment, masked arrays, gradient batches and replayed simulation.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Test the end-to-end beginner handoff without revealing reference answers.
Resume and profile a sharded training run
In developmentLogical CPU ownership, weighted updates, recovery and trace evidence.
Evidence: Reference check pending for this source. Teaching review pending.
Help needed: Qualify fresh-process multi-host data and recovery behavior on an explicitly named target.
Build a causal Transformer training and inference harness
Available draftTiny causal training, recovery, KV cache and export, plus a separate train-only BPE decoder workflow with natural chapter splits, measured short-run results and an explained baseline comparison.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Broaden source and tokenizer studies, complete larger staged training and assess generation on genuinely unseen material. Repeatedly inspected Alice chapters are not a blind benchmark.
Audit conditioning, gradient noise and optimizer decisions
Available draftGeometry, conditioning, enumerated gradient noise and optimizer-state evidence.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Pilot the synthesis with learner implementations and review failure explanations.
Ship a tracked and containerized model release
Available draftTracked training, slice gates and deterministic LLM application fixtures.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Evaluate a real application with versioned prompts/retrieval, adversarial cases and human review.
Audit pretraining and post-training objectives
Available draftEight small masked, contrastive, adaptation and preference objectives.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Connect each objective to real saved architectures and phase-to-phase checkpoint lineage.
Convert PyTorch weights and audit Flax architecture parity
Available draftReal framework dense mappings and first-divergence numerical checks.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Convert a complete saved attention architecture, including normalization, masks and input gradients.
Check the CPU environment and a saved matrix arithmetic identity
Available draftEnvironment evidence and independent arithmetic with learner-owned implementation.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Run an unaided novice walkthrough and review the retained environment record.
Audit masked batch moments and padding-sensitive denominators
Available draftArray shape, statistics and changed-input diagnostics.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add ragged/masked data and assess the learner explanation of each axis.
Compare scalar, vmap, autodiff, and changed-shape JIT behavior
Available draftDifferentiation, batching and tracing invariants.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add composed transformations with a deliberately incorrect derivative boundary.
Resume a complete training stream across a process failure
In developmentLocal checkpoint contract and independent continuation checks.
Evidence: Reference check pending for this source. Teaching review pending.
Help needed: Inject incomplete writes, stale manifests and incompatible input state in separate processes.
Measure CPU JIT latency and changed-shape traces with retained JAXPR evidence
Available draftCompilation/timing distinctions and retained evidence.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add a target-backed workload and a reviewed trace, not a generic timing improvement claim.
Audit SFT, DPO and sequence PPO
In developmentBounded adaptation and preference-contract checks.
Evidence: Reference check pending for this source. Teaching review pending.
Help needed: Transfer the contracts to a saved Transformer and document retained behavior.
Inspect CPU device inventory, BF16 storage, and collective participant arithmetic
Available draftCPU-verifiable runtime and participation evidence contracts.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Observe a learner-owned TPU receipt on a declared topology and keep the exact numerical reference.
Gate simulated multi-worker checkpoint publication on complete receipts
Available draftLaunch, checkpoint and recovery evidence contracts.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Add a fresh-process multi-host restart and failure-injection qualification.
Estimate optimizer-state bytes and calculate supplied-rate scaling efficiency
Available draftPrecision/memory/profiling contracts with bounded assumptions.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Attach actual native low-bit execution and an interpreted target trace.
Reproducible JAX/Flax experiments workshop
Available draftSix recorded CPU runs, paired comparisons, logging and a research notebook.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Pilot learner-owned hypotheses, failed/retried runs and the optional logger adapters.
Trace a tiny bidirectional encoder from pretraining to retrieval
In developmentA real-text CPU pilot with MLM, contrastive continuation, adaptation and held-out retrieval; quality is weak.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Improve the frozen data/evaluation recipe against BM25 before scaling or claiming a useful released embedding model.
Trace a tiny causal Transformer from pretraining to instruction-style adaptation
In developmentA real-text CPU Transformer pilot with continuation, held-out perplexity and export; quality is weak.
Evidence: CPU reference checker passed for current sources. Teaching review pending.
Help needed: Improve baseline-relative held-out quality and tokenizer/data decisions before scaling or publishing a production model.
Planned model capstones
Train a CLIP image–text model
PlannedLearn paired representations, test zero-shot classification, and qualify both retrieval directions before releasing the encoders.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Train and compare a SigLIP model
PlannedReplace the softmax objective with pairwise sigmoid training. Audit labels, reduction, negatives and distributed gradients against a controlled CLIP baseline.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build a measured retrieval service
PlannedTake a qualified text or cross-modal encoder through exact search, approximate indexes, hybrid retrieval, reranking, evaluation and safe index migration.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build a staged embedding model
PlannedMLM to contrastive pretraining and retrieval fine-tuning, with frozen data, MTEB evaluation and release gates.
Evidence: Bounded CPU pilot evidence verified; full capstone remains planned; evidence index projects/embedding-model-300m/reference/evidence-index.json (SHA-256 011b027f03dc7bb7913ae00b48864b92d9c09c378d6d07387739083cdce40ea3). Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build a causal language model
PlannedConnect tokenizer/data preparation, next-token pretraining, adaptation, generation evaluation and model publication.
Evidence: Bounded CPU pilot evidence verified; full capstone remains planned; evidence index projects/causal-model-300m/reference/evidence-index.json (SHA-256 e3170945eada0dfb732e8acf3ee09a7d02399983325f623e0082625c6b17ec78). Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build an image understanding model
PlannedCarry a frozen image recipe through representation learning, task adaptation and measured evaluation.
Evidence: Bounded CPU pilot evidence verified; full capstone remains planned; evidence index projects/image-model-300m/reference/evidence/v14/packet-manifest.json (SHA-256 3b1dcfe589fffc25a43da336beb07d82334b6f8971a73084e2b989b6e8f2a794). Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build an image generation pipeline
PlannedQualify image representations, denoising training, conditioning and sampling with separate quality and runtime evidence.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build an audio understanding model
PlannedVersion waveform and acoustic representations through training, recognition or event adaptation and robustness evaluation.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build an audio generation pipeline
PlannedConnect codec qualification, sequence or denoising training, conditioning and waveform evaluation.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Build a grounded multimodal model
PlannedQualify encoders, connector alignment, instruction adaptation and grounded evaluation, keeping retrieval a distinct outcome.
Evidence: Specification only; no runnable capstone. Implementation and review pending.
Help needed: Implement the scoped data recipe, architecture, training stages, independent evaluation and exported reload before attempting the target scale.
Synthesis assessments
Regression synthesis
Available draftDefend a JAX regression experiment with independent gradient checks, reproducible training, learning-rate analysis and a repair for a broadcasting bug.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Classifier and image lifecycle synthesis
Available draftAssess a JAX image classifier through stable losses, preprocessing, checkpoint recovery, precision tradeoffs and verified exported inference.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Math & optimization synthesis
Available draftExplain JAX optimizer behavior using least-squares geometry, Hessian eigenvalues, minibatch gradient noise, clipping and independent Adam update checks.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Scientific inverse-problem synthesis
Available draftDefend a differentiable JAX inverse problem with conservation checks, sensitivity analysis, parameter recovery and explained numerical error.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Reinforcement-learning synthesis
Available draftAssess JAX reinforcement learning with episode-boundary tests, independent policy expectations, clipped updates and held-out evaluation.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Bayesian regression synthesis
Available draftCompare exact Bayesian regression with HMC and variational inference in JAX, then diagnose approximation error and posterior predictive behavior.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Derivative synthesis
Available draftDefend JAX derivative transformations with rectangular Jacobians, custom VJP rules, primitive extensions and independent numerical comparisons.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Shipping synthesis
Available draftJustify a JAX model release with adaptation comparisons, artifact checks, quantized inference measurements and transparent service capacity estimates.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Workload operations synthesis
Available draftRehearse JAX workload incidents with process outcome checks, complete-state recovery, model validation, rollback drills and capacity reasoning.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Profiling, partitioning and kernel synthesis
Available draftEvaluate a sharded JAX workload with independent numerical checks, profiler evidence, partitioning analysis and a bounded Pallas kernel decision.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Training-system synthesis
Available draftAssess a JAX causal training system through checkpoint recovery, cache precision, exported inference and a separate TPU runtime qualification plan.
Evidence: Written tasks, public scaffolds and rubric; no calibrated grading study. Rubric and learner pilot review pending.
Help needed: Trial the tasks with unseen learner implementations and record rubric disagreements, changed cases and reviewer reasoning.
Learning paths
Start with JAX
Available draftLearn how arrays, derivatives and explicit state fit together by building and checking a regression experiment. Use the deeper math lessons when you need to explain why an update works or fails.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Teaching review is still pending.
Train neural networks
In developmentTrain a CPU image classifier, preserve its complete state through interruption, inspect held-out errors and verify its exported inference. Language modeling and TPU work are optional branches.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: deployment.
Simulate and discover
In developmentSimulate physical dynamics, check numerical and derivative accuracy, and recover unknown parameters with held-out evidence. Build a reusable CPU inverse-problem audit before extending to adaptive solvers or accelerators.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: science.
Model uncertainty
In developmentExpress assumptions as probability models, verify a known posterior, then implement HMC and variational inference with explicit diagnostics. Use predictive checks to distinguish a sampler problem from a model that does not explain the data.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: probability.
Learn by interacting
In developmentTrain a policy from actual environment interaction, verify its objective against an independent small-state reference, and compare multiple seeded agents on fixed held-out evaluation streams.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: rl.
Make it fast. Make it scale.
In developmentProfile a correct workload, preserve its global objective while partitioning arrays, and audit actual Pallas kernels before requesting target measurements. CPU semantics are available; accelerator speed and multi-host claims require their own receipts.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: distributed, kernels.
Understand JAX itself
Available draftTrace a function into jaxpr, verify forward and reverse sensitivities, write exact custom derivative rules, and compile a small transformation whose limits you can explain.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Teaching review is still pending.
Learn TPUs and Google Cloud
In developmentBuild and recover an experiment on CPU, then learn GCP provisioning, TPU generations, training and inference memory, precision policies and XProf profiling. Connect measured error and real target traces to model engineering.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: tpu, tpu-workflows, tpu-systems, transformers, pretraining, rl, posttraining, distributed, deployment, operations.
Ship and adapt JAX workloads
In developmentMove trained computations into a measured inference service. Check framework/precision parity, package a non-root Docker image, verify readiness and connect evaluation, model identity and rollback evidence.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: deployment.
Operate JAX workloads
In developmentOperate reproducible model workloads from tracked experiments to containerized inference. Practice failure recovery, MLOps release gates, LLMOps evaluation and tracing, and ModelOps ownership, approval and rollback.
Evidence: Linked practice and capstone references; route comprehension unvalidated. Teaching review pending.
Help needed: Pilot this complete route, verify its prerequisite effort and review the final learner evidence. Parts still in development: distributed, deployment, operations.
Career guides
ML & training engineer
Available draftTurn a dataset and model idea into a reproducible experiment. You define the objective, diagnose learning, keep held-out evaluation isolated and prove a saved run can continue correctly.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
Research engineer
Available draftTurn a proposed method into an experiment that can support or refute a claim. Your work connects mathematical assumptions, a simple baseline, independent checks and carefully bounded conclusions.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
Scientific ML engineer
Available draftConnect observations to a numerical or physical model. You check derivatives and solver behavior, fit parameters and explain how noise, conditioning and model assumptions affect the answer.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
RL engineer
Available draftBuild experiments in which actions change later observations. You keep environment state, random sampling and policy updates reproducible, and evaluate improvement across seeds and conditions.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
Accelerator & performance engineer
Available draftMeasure where a workload spends time and memory, then improve the limiting operation while preserving its numerical contract. Explain when an optimization helps and when it does not.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
Inference & deployment engineer
Available draftPackage a trained model so another process or device can use it reliably. Own preprocessing, exported predictions, precision choices and measurements under a defined request workload.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
JAX / accelerator operations engineer
Available draftMake model workloads observable and recoverable. Track progress and resource use, diagnose failed or stalled runs and rehearse recovery before relying on it.
Evidence: Internal capability map; no external role or hiring validation. Practitioner review pending.
Help needed: Practitioners: review the responsibilities, missing skills, evidence criteria and prerequisite burden for this role.
Modality tracks
Image: from pixels to a deployed classifier
In developmentA connected CPU image harness is runnable: a trained CNN, full fresh-process recovery, interpreted clean/corrupted evaluation, calibrated integer arithmetic, serialized inference and measured rollout/rollback. Local image ingestion is implemented; real-photo quality and target-device performance remain unverified.
Evidence: Bounded harness practice; full modality capstones planned. Teaching review pending.
Help needed: Connect the runnable harness to licensed natural data, meaningful task benchmarks and a separately qualified deployment target.
Text: from tokens to reproducible generation
In developmentRunnable four-stage CPU byte-Transformer harness: real training, complete fresh-process recovery, fixed held-out evaluation, FP32/BF16/INT8 KV caches, six inference exports and profiler traces. Includes a local document loader. External-corpus quality, service load and physical TPU qualification remain extensions. An additional tiny bidirectional embedding pilot connects three objective phases and reports held-out retrieval against BM25, including a poor measured outcome. A separate 70,704-parameter causal pilot connects real-text next-token pretraining, response-masked continuation adaptation, chapter-held-out perplexity against a unigram baseline, and verified export sampling. Its weak evaluation and repetitive sample are retained.
Evidence: Bounded harness practice; full modality capstones planned. Teaching review pending.
Help needed: Connect the runnable harness to licensed natural data, meaningful task benchmarks and a separately qualified deployment target.
Audio: from waveforms to an edge sound classifier
In developmentA complete CPU waveform-to-inference harness is runnable: trained spectral classifier, full fresh-process recovery, fixed evaluation, calibrated precision and complete exported artifacts. PCM WAV ingestion is available; natural-recording quality and target-edge performance remain unmeasured.
Evidence: Bounded harness practice; full modality capstones planned. Teaching review pending.
Help needed: Connect the runnable harness to licensed natural data, meaningful task benchmarks and a separately qualified deployment target.
Cross-modal: connect images and text
In developmentA connected CPU image–text retrieval harness is runnable: both encoders train, complete state restores in a fresh process, calibrated precision policies export and reload, and a measured release gate supports rollback. The synthetic four-concept fixture is verified separately from natural-language or real-image retrieval quality.
Evidence: Bounded harness practice; full modality capstones planned. Teaching review pending.
Help needed: Connect the runnable harness to licensed natural data, meaningful task benchmarks and a separately qualified deployment target.
Practical guides
Tools, folders and experiment organization
Available draftA practical guide with a downloadable practice workspace.
Evidence: Workspace checks available; novice review pending. Teaching review pending.
Help needed: Pilot the setup and evidence handoff with a new learner; review platform-specific and hardware claims.
TPU and Google Cloud workflow
Available draftA practical guide with a downloadable practice workspace.
Evidence: Workspace checks available; novice review pending. Teaching review pending.
Help needed: Pilot the setup and evidence handoff with a new learner; review platform-specific and hardware claims.
TPU precision and profiling
Available draftA practical guide with a downloadable practice workspace.
Evidence: Workspace checks available; novice review pending. Teaching review pending.
Help needed: Pilot the setup and evidence handoff with a new learner; review platform-specific and hardware claims.
Book formats
JAX Pathways book
In developmentHTML and EPUB include all authored chapters, figures, practical appendices and solutions.
Evidence: Automated export checks; native reader and teaching qualification incomplete. Teaching review pending.
Help needed: Review math/code on native EPUB readers, print/PDF and assistive technology; contribute chapter-level teaching reviews.