| language: | |
| - en | |
| tags: | |
| - writing-feedback | |
| # writing-feedback | |
| E1 现有反馈、原 prompt、离线数据和冻结训练/测试划分 | |
| Public research asset export approved by the owner. | |
| ## Scope and provenance | |
| Files preserve the existing experiment assets. Their filenames and hashes are recorded in `asset_manifest.json`. | |
| GPU environment validation is separate and has not been completed. | |
| Original JSONL/Parquet formats are preserved. Test data and training data are explicitly separated by directory; do not combine them for training. | |
| WritingPrompts is the historical prompt source; LitBench is the independent evaluator. The critic export retains four dimension scores and verbal feedback. When deriving new pass/fail or score-conditioned training data, apply the supplied test-prompt exclusion manifest. Raw critic records include held-out prompts and are not a ready-to-train split. | |
| This export does not assert new licensing or provenance guarantees beyond the underlying sources. See the original source terms before redistribution. | |