Milestone 7 of 8
Save and replay the toolkit
Write configuration, vocabulary, token records, encoded sequences, unknown traces, fixtures, source identity, and a manifest, then replay separately.
The toolkit should be usable from its saved configuration, vocabulary, records, and identity. Replay it separately before treating the artifacts as durable.
Goal
Write configuration, vocabulary, document token records, encoded sequences, unknown traces, summaries, fixtures, source identity, figures, and a manifest. Replay the toolkit in a separate directory and record agreement or mismatch.
Inputs
Use all artifacts from the earlier milestones and the original inspected corpus records. The manifest must identify the corpus, normalization and scanner versions, reserved IDs, vocabulary order and size, boundary policy, artifact paths, relevant runtime versions, and implementation revision.
Generate token-frequency-rank and per-document token/unknown-count figures from
saved records and preserve their exact plot data. The replay starts from saved
configuration and source identity under a separate directory such as
output/replay/.
Deliverables
Extend src/main.py and src/report.py, or add a clearly separated replay
function, to save:
output/tokenizer_config.json;output/vocabulary.jsonand the escaped table;- per-document canonical and token records;
- encoded sequences and
output/unknown_tokens.csv; - document and corpus summaries;
- exact normalization, scanning, vocabulary-order, codec, and round-trip fixtures;
- figures and their plot data;
output/tokenizer_manifest.json; and- a replay agreement or mismatch record under
output/replay/.
The manifest should reference artifacts rather than copy every sequence into one record. Never overwrite the original output during replay.
Checks
Read every required artifact back and check identity, schema, configured order, contiguous IDs, span coverage, unknown traces, summaries, fixtures, and plot data. Replay an unchanged toolkit separately and compare each transformation and artifact. Confirm the original output remains unchanged.
Deliberately change normalization version, reserved order, one vocabulary ID, one token span, or one unknown trace and check that replay identifies the first difference. An incompatible identity or runtime must not be reported as exact agreement.
Workspace
Keep original artifacts under output/ and replay artifacts under a distinct
child directory. Keep output/tokenizer_manifest.json and project-level
report.md at stable paths.
Hints
HintReplay transformations in order
HintPlot data is an artifact
Review
Open original and replay artifacts side by side. Explain what agreement establishes for this corpus and configuration, and why it does not establish that the tokenizer is best or beneficial to a model.
How to check your work
Checks compare manifest and replay evidence with the version-matched reference fixture. The supplied fixture makes any changed tokenizer identity explicit.