Do this only if the format matches
Confirm the round with recruiting first. This is an original practice exercise, not a claim that a specific company uses this exact prompt.
Protocol
- Run contract tests and classify failures as data, model, update, or evaluation defects.
- Trace one token through input, target, mask, logits, loss, and gradient.
- Repair label and causal-mask semantics before changing the optimizer loop.
- Make accumulation, clipping, and scheduler cadence agree on what an update means.
- Restore evaluation isolation and caller mode.
- Overfit one tiny batch, then record the first remaining multi-GPU risk.
Deliverables
- Repaired code and one new regression test.
- A defect table with symptom, mechanism, and verification.
- A one-batch overfit result.
- A prioritized plan for a distributed recurrence.
Graduation gates
Targets are next tokens and padded targets are ignored.
Prefix outputs cannot depend on future tokens.
Cross-entropy receives raw logits.
One accumulation window creates one optimizer and scheduler step.
Starter files
Assignment READMEDiagnostic order and deliverables.Open file →Data contractContains label defects.Open file →Tiny causal LMContains attention and logits defects.Open file →Training protocolContains update and evaluation defects.Open file →Public testsCovers only part of the failure surface.Open file →