# Reconstructing the historical CoEdIT inputs

`specification.json` pins the official upstream files and maps reference recipes
to local output paths. IDs are source-distribution IDs; SHA256 checks bind each
reference to its exact source or target string after instruction-prefix removal.
`manifest-*.jsonl` and `clean-shard-*.jsonl` contain metadata and references only.
`corruptions.jsonl` stores source references, generated additions and character
edit programs. A `slice` is a half-open character interval into the separately
obtained source string. In inverse-edit payloads, slices refer to the reconstructed
corrupted string. These recipes do not include full sentence text or token IDs
that encode the original sentences.

Run `reconstruct_coedit.py` as described in the root README. It recreates the
training/evaluation pairs, identity inputs, clean target shards and approved
corruption training fields. It does not re-run GPT generation or promise the
same Parquet bytes. Raw API traces and model-operation prose are not reconstructed.

## Optional fine-tuning corpus serialization

With the CES code environment and canonical tokenizer already installed, set
`NANOCHAT_BASE_DIR` to a runtime directory containing `tokenizer/`. From the CES
code checkout, use the reconstructed manifests:

```bash
uv run ces prepare corpus \
  --train-manifest ../ces-local/data/evaluations/coedit-gec-transfer-v1/train.jsonl \
  --validation-manifest ../ces-local/data/evaluations/coedit-gec-v1/development.jsonl \
  --output-dir ../ces-local/data/finetuning/coedit-gec-transfer-corpus \
  --train-shards 8 --seed 20260727
```

This recreates the eight training shards and development shard in `raw/` and
`candidates/`. Canonical character and token edit programs are derived locally.
The tokenizer is an external prerequisite for this optional step; it is not
included in the data-only release. The text reconstruction step needs no tokenizer.