遇见数据集

qodex-teacher-codex-r2-clean-48k

收藏
魔搭社区2026-10-31 更新2026-11-01 收录
官方服务:

资源简介:

# Qodex Teacher Codex R2 Clean 48k Cleaned merged Codex CLI teacher trajectory dataset rendered as Qwen ChatML with a 48k token budget. ## Files - `data/train.chatml.jsonl`: 1146 cleaned ChatML rows. - `data/train.prompt_completion.jsonl`: prompt/completion expansion for SFT tooling. - `manifest.json`: clean export manifest. - `reports/row-manifest.csv`: row-level source provenance. - `reports/dropped-rows.json`: rows removed during cleaning. - `reports/preclean-*.json`: pre-clean duplicate/leak checks. - `CHECKSUMS.sha256`: SHA-256 checksums for primary artifacts. ## Source Canonical local source: `exports/20260503T024220Z-qodex-teacher-v0.3-clean-48k/chatml/qwen_chatml_train_48k_v3_clean.jsonl` The clean ChatML SHA-256 is: `8fde77003c24b8170b98f668121aaf1db446e02bdc87f07b3f0ff63d8e3e6d92`. The source was built from the 528-row teacher-suite-v2 pool plus qodex-v0.2 rows, then filtered for missing web-search tool-call IDs and one known verifier-gaming row. See `manifest.json` for exact counts.

提供机构:
maas
创建时间:
2026-06-22
二维码
社区交流群
二维码
科研交流群
商业服务