Attribution and changes
Louis Robinson, Causal Edit Serialization (CES), 2026: https://github.com/Louiii/CES.
NVIDIA, Nemotron-ClimbMix: https://huggingface.co/datasets/nvidia/Nemotron-ClimbMix; associated paper https://arxiv.org/abs/2504.13161. Source text was obtained through Andrej Karpathy's shuffled distribution: https://huggingface.co/datasets/karpathy/climbmix-400b-shuffle. CES applies filtering, synthetic corruption, edit annotation, fixed partitioning and serialization. The original distribution's CC BY-NC 4.0 statement is preserved.
Vipul Raheja, Dhruv Kumar, Ryan Koo and Dongyeop Kang. CoEdIT: Text Editing by Task-Specific Instruction Tuning, Findings of EMNLP 2023: https://aclanthology.org/2023.findings-emnlp.350/. Official distribution: https://huggingface.co/datasets/grammarly/coedit. CES's reconstruction removes instruction prefixes, uses fixed deduplicated partitions, and applies recorded synthetic edits locally. This repository contains references and edit recipes, not the original CoEdIT sentence pairs.
CoEdIT credits NUCLE (Daniel Dahlmeier, Hwee Tou Ng and Siew Mei Wu), NAIST Lang-8 (Toshikazu Tajiri, Mamoru Komachi and Yuji Matsumoto), and W&I–LOCNESS/BEA-2019 (Christopher Bryant and collaborators) as GEC sources. Corpus terms remain applicable to separately obtained data; the CoEdIT card's Apache 2.0 statement is not presented here as overriding those terms.
GPT-5 Mini generated synthetic corruptions. They are research annotations, not human-verified ground truth. API transport records are excluded. Five pretraining candidate rows with credential-shaped strings were omitted whole, preserving text/offset alignment in retained examples. Arithmetic experiments are excluded.
The CES corpus-building code extends Andrej Karpathy's NanoChat; the code repository preserves its MIT copyright notice.
The CES checkpoints were trained using ClimbMix-based data, with CoEdIT-derived fine-tuning/post-training where specified by their adjacent metadata. The weights and architecture changes are original CES contributions; source-data credits and restrictions remain as described above.