| tags: | |
| - eklav | |
| - passagereranking | |
| - answer-only | |
| # Eklav-Reranker-AnswerOnly-Data | |
| Training data for the **Eklav** paper. | |
| - **Task:** passage reranking (BRIGHT / NevIR benchmarks) | |
| - **Method:** Answer-only (no reasoning of any kind -- the no-CoT floor) | |
| - **Examples:** 381,934 train / 3,857 held-out val | |
| - **Format:** ShareGPT (`system` + `conversations: [{from, value}]`), used for LoRA SFT via LLaMA-Factory. | |
| Single-turn ShareGPT conversations. Each row: a query+passage relevance-judgment prompt (human turn) and a bare `true`/`false` judgment (gpt turn) -- no hint, no chain-of-thought, no `<think>` block anywhere. Loss is computed over the entire (one-token) gpt turn. | |
| Built from the same source query/passage/judgment triples as `Eklav-Reranker-Data` (Eklav method) and `Eklav-Reranker-CotGen-Data` (std-SFT method), with the teacher reasoning removed entirely rather than hinted at or reproduced -- the train/val split (381,934 / 3,857, seed 42) matches those two repos exactly so all three methods are directly comparable example-for-example. | |
| Files: | |
| - `train.json` -- training split | |
| - `val.json` -- held-out validation split | |
| See the Eklav paper for full dataset construction methodology, and the corresponding `Eklav-*-PassageReranking-AnswerOnly` model repos for checkpoints trained on this data. | |