Datasets:
| license: unknown | |
| task_categories: | |
| - text-generation | |
| tags: | |
| - preference | |
| - dpo | |
| - reward-model | |
| size_categories: | |
| - 1K<n<10K | |
| # TEXT_CLASSIFICATION | |
| A preference dataset for **TEXT_CLASSIFICATION**, harvested from real, human-labelled sources and curated by an automated harvesting harness with an LLM quality gate. | |
| ## Format | |
| Standard preference / DPO schema — each row: | |
| | column | meaning | | |
| |---|---| | |
| | `prompt` | the request (originally `text`) | | |
| | `chosen` | the human-preferred response | | |
| | `rejected` | a worse response to the same prompt | | |
| | `source` | the dataset/URL the row was harvested from | | |
| ## Splits | |
| 80/10/10 `train` / `validation` / `test` (seeded shuffle): train:1299 / validation:162 / test:163 | |
| ## Stats | |
| - **Rows:** 1624 | |
| - **Distinct sources:** 4 | |
| ### Sources | |
| - `clinc/oos-eval (data_full.json)` | |
| - `fancyzhx/ag_news` | |
| - `mteb/banking77` | |
| - `synthetic:deepseek/deepseek-v4-flash-0731` | |
| ## Provenance | |
| **`prompt` and `chosen` are real** — harvested from the human-labelled sources listed above (accepted / upvoted / high-rated responses). | |
| **`rejected` is partly synthetic.** Most sources retain only the accepted answer, so where no genuine low-scored counterpart existed the negative was generated by defect injection: an LLM rewrites that row's own `chosen` with exactly one flaw introduced (omission, factual error, wrong remedy, unsupported claim, or misprioritisation) at matched length. Rows carrying a `defect` column are the generated ones; rows without it kept a negative that was present in the source. | |
| Rows passed a per-row check (negative is unique, not a stub, not spliced from `chosen`) and a whole-corpus gate for repeated negatives, length shortcuts, train/eval leakage and bag-of-words separability. | |