README.md

---
pretty_name: Causal Edit Serialization (CES) — Data and checkpoints
license: other
license_name: ces-component-licenses
license_link: https://huggingface.co/datasets/lorob/CES/blob/main/LICENSES.md
language:
- en
tags:
- causal-edit-serialization
- text-editing
- grammatical-error-correction
configs:
- config_name: generated-edit-candidates
  data_files:
  - split: train
    path: data/pretraining/gpt5mini-edit-candidates/*.parquet
- config_name: controlled-repair
  data_files:
  - split: validation
    path: data/evaluations/controlled-edit-val-v1/manifest.jsonl
- config_name: balanced-controlled-training
  data_files:
  - split: train
    path: data/posttraining/ces/v1/balanced-sft/controlled-edit-train-balanced-v1.jsonl
---

# Causal Edit Serialization: data and checkpoints

Louis Robinson · [CES code and paper project](https://github.com/Louiii/CES)

This repository contains CES training candidates, controlled-repair evaluation
data, and recipes for reconstructing the CoEdIT-based inputs locally. It also includes 66 model checkpoints with their configuration metadata.
Experiment logs are excluded.

| Directory | Contents |
|---|---|
| `data/pretraining/` | 361,768 synthetic-edit candidate rows in 109 Parquet files, plus copy-scan records |
| `data/evaluations/controlled-edit-val-v1/` | 1,000 fixed controlled-repair examples |
| `data/posttraining/ces/v1/balanced-sft/` | 1,500 controlled recovery training examples |
| `reconstruction/` | CoEdIT example references, hashes, generated edit recipes and a local reconstruction script |

The full raw ClimbMix corpus is not included. Its source passages occur in the
prepared ClimbMix-based examples and retain the original CC BY-NC 4.0 terms.
Five candidate rows containing credential-shaped strings were omitted whole;
retained text and edit offsets are unchanged. The released candidate subset is
therefore not byte-identical to the historical training input. Arithmetic
experiments are excluded.

## Why the CoEdIT sentences are not included

We would prefer to distribute all prepared inputs directly, so reproducing CES
would require fewer setup steps. However, the [CoEdIT dataset card](https://huggingface.co/datasets/grammarly/coedit)
advertises Apache 2.0 while [its paper](https://aclanthology.org/2023.findings-emnlp.350/)
identifies underlying GEC corpora with additional restrictions. In particular,
[NUCLE's source agreement](https://www.comp.nus.edu.sg/~nlp/conll14st/nucle_license.pdf)
restricts redistribution, and [NAIST Lang-8](https://sites.google.com/site/naistlang8corpora)
limits availability to research and educational purposes. We have not established
how the distribution's Apache statement resolves those underlying permissions.

To avoid asserting rights we cannot establish, this release provides references
and preparation code instead of republishing the CoEdIT source/target sentences,
their serialized copies, or complete generated corruptions. This adds a download
and reconstruction step. It is a limitation of the release, not a claim that the
CoEdIT authors refused us permission or acted improperly.

Users obtain the source files from the official CoEdIT distribution and remain
responsible for complying with the terms applicable to their use. Reconstruction
does not remove the original restrictions or create new permissions.

## Reconstruct the CoEdIT inputs

Install `huggingface_hub` and `pyarrow` in a Python environment. From this
repository's downloaded directory, run:

```bash
python reconstruction/reconstruct_coedit.py \
  --download \
  --upstream-dir ../ces-upstream/coedit \
  --output ../ces-local
```

The download uses the official `grammarly/coedit` repository at commit
`e9a255c33ef910bc33a9d2b522653fa87521583e`. This revision has been verified against
the retained CES examples; it is not claimed to be the unrecorded historical
download revision. File and text hashes are checked before reconstruction.
To use source files you already obtained, omit `--download` and place the pinned
`train.jsonl` and `validation.jsonl` files in `--upstream-dir`.

The script reconstructs, in historical order:

- 12,756 training, 485 development, 2,000 selection and 5,000 confirmation pairs;
- the 5,000 clean confirmation inputs and 12,000 clean corruption-generation targets;
- the 6,813 approved generated corruptions, including the historical character
  edit programs used by the training loader.

Generated corruption recipes contain offsets into source text and added payloads,
not complete source sentences. No model/API calls are needed to reconstruct them.
The training fields match the historical records; API response traces and raw
model-operation prose are omitted. Parquet binary files need not be byte-identical.

Keep upstream downloads and reconstructed outputs outside this public repository.
The script enforces separate directories to reduce accidental re-upload.
See [reconstruction/README.md](reconstruction/README.md) for rebuilding the
optional tokenizer-dependent fine-tuning shards using CES code.

## Download only what you need

```python
from huggingface_hub import HfApi, snapshot_download

revision = HfApi().repo_info("lorob/CES", repo_type="dataset").sha
print("Save this artifact revision:", revision)
snapshot_download(
    "lorob/CES", repo_type="dataset", revision=revision,
    local_dir="ces-data",
    allow_patterns=["reconstruction/*", "LICENSES.md", "NOTICE.md", "licenses/*"],
)
```

Use `data/pretraining/*`, `data/evaluations/*` or `data/posttraining/*` in
`allow_patterns` for those components. `SHA256SUMS` covers every released file
except itself and `manifest.jsonl`; verify a complete download with
`sha256sum -c SHA256SUMS` from its directory.

## Licences and attribution

CES reconstruction code is MIT; original annotations, edit recipes and
documentation are CC BY 4.0, limited to rights held by the author. ClimbMix source
text retains CC BY-NC 4.0. These are component-specific permissions, not a blanket
unrestricted licence. See [LICENSES.md](LICENSES.md) and [NOTICE.md](NOTICE.md).

## Checkpoints

`checkpoints/` contains 48 pretrained/CoEdIT-finetuned method-grid weights,
9 final CES RL weights, 6 matched clean-AR references, and 2 pointer-ablation
weights. Each `model_*.pt` has its adjacent `meta_*.json`. Total size is approximately
132.9 GiB; use selective downloads for the model you need. No optimizer states or
training logs are included.

These custom checkpoints require CES code and the canonical `ces32k_v1` tokenizer;
they are not Transformers `AutoModel` repositories. The tokenizer is a separate
prerequisite and is not included in this checkpoint upload. For example, select
`checkpoints/rl-posttraining/ces/d12/posttrain-seed-42/*` in the download example
above for the smallest final CES RL model.

Original CES weights are CC BY 4.0 to the extent of the author's rights; this does
not relicense training text or establish permissions for all downstream uses.

### Checkpoint identities

`checkpoints/models.json` lists model IDs, download paths and complete training lineages. The d12 RL family uses a separate pretrained base at `checkpoints/rl-posttraining/ces/d12/pretrained/model_074800.pt`; the d12 methods-grid pretrained checkpoint is a different run. Adapted-CES includes clean pretraining and CES adaptation within one experiment; its internal clean-stage weights are not distributed.