Milestone 8 of 8
Write and audit the dataset card
State source identity, example rules, shapes, groups, batches, padding, unknowns, empty documents, replay evidence, and unsupported claims.
The dataset card is a short account of what the saved examples contain and what they do not establish.
Goal
Write and audit report.md so another reader can reproduce the sequence and
understand its identities, padding, groups, batches, unknowns, remainders, and
limits.
Inputs
Read the saved configuration, group assignment identity, examples metadata, per-group arrays, summaries, batch manifest, selected-example records, figures and exact plot data, sequence manifest, and replay record. Do not rely on current editor output or restate a count that is not in an artifact.
Deliverables
Complete the project-level report.md with these sections:
- Identity and scope: name corpus, tokenizer, configuration, group assignment, document order, and the bounded fixture.
- Example rule: explain preceding contexts, target positions
1throughlen(ids)-1, context width,<bos>,<eos>, empty documents, and no cross-document history. - Padding and arrays: state shapes, dtypes, left-padding, mask meaning, real-context lengths, and the mask-sum invariant.
- Groups and batches: state supplied group labels, shared indexes, order, kept remainders, and empty-group behavior.
- Evidence and limits: link artifacts, selected rows, figures, replay, unknown behavior, and at least one concrete limit on interpretation.
Audit one known row and one <unk> row against saved records. Make clear that
the group names are not a training/validation/test claim and that one replay
agreement does not establish independence, balance, learning, or model
quality.
Checks
Check every required rule and fixed field appears in the card and agrees with the manifest and fixtures. Confirm that example counts conserve per-document counts, group totals, and batch totals; that remainders and empty groups are named; and that unknown tokens remain visible.
Reject unsupported claims about language modeling, independence, balanced groups, leakage, usefulness, or downstream performance. Replace praise with an artifact-backed observation or remove it.
Workspace
Keep the card at report.md. Do not duplicate full arrays or metadata in
prose. The saved files remain the primary evidence.
Hints
HintFollow one row
HintName the boundary
Review
Read the card as a later language-model project author. Can they identify the exact context and target rule, reproduce the arrays, recognize the fixed group labels, keep the final batch, and understand what the fixture cannot prove?
How to check your work
Checks compare the card's sections, counts, identities, limits, and replay statement with the supplied fixture. The supplied fixture is an evidence template, not a result to copy.