SAFWAAN's picture
Upload README.md
b8048f1 verified
|
Raw History Blame Contribute Delete
5.76 kB
---
license: cc-by-sa-4.0
language:
- en
task_categories:
- question-answering
- text-generation
tags:
- rag
- grounded-qa
- citations
- abstention
- field-atlas
size_categories:
- 1K<n<10K
---
# Field Atlas reader training data
9,497 training examples and a 400-item held-out set for teaching a small model to answer from given passages with citations, to say **"The saved sources don't say."** when they do not answer, and to name the missing part when they only half answer. Built for [poidh bounty #45](https://poidh.xyz/mainnet/bounty/45) and used to train [SAFWAAN/fieldatlas-reader-qwen3.5-2b](https://huggingface.co/SAFWAAN/fieldatlas-reader-qwen3.5-2b).
## Files
| File | Rows | What it is |
|---|---|---|
| `train_items.jsonl` | 9,497 | Full training items in the benchmark's schema: `id`, `category`, `question`, `passages` (`n`, `title`, `text`, `chunkId`, `source`), `answerable` (`yes` / `partial` / `no`), `reference` (the target answer), `keyFacts`, `cites`, `pack` |
| `train_chat.jsonl` | 9,497 | The same items as chat messages (`system`, `user`, `assistant`), ready for SFT. The system and user texts are produced by `reader_bench.SYSTEM` and `reader_bench.prompt()` from the Field Atlas repo, so they are identical to what the app sends |
| `heldout.jsonl` | 400 | Same schema as `train_items.jsonl`; scored with `reader_bench.py score`. Frozen before any training; never used for model selection |
Label mix (training): about 41 % fully answerable, 33 % partly answerable, 26 % not answerable. Categories: fact, explain, travel, compare, biology, synthesis, crypto, health, first aid, insufficient-information. Passages per item: 1–8, mostly 4–7, like the benchmark.
## How each example was made
The pipeline (`gen.py`, `fetch.py`, `run.sh` in the model repo) mirrors how the benchmark's own items were built:
1. **Seed passage** drawn from the Field Atlas knowledge packs the app actually installs (`fetch.py` downloads the pinned `.fapack` files from [0x94t3z/fieldatlas-knowledge](https://huggingface.co/datasets/0x94t3z/fieldatlas-knowledge) and [0x94t3z/fieldatlas-encyclopedia](https://huggingface.co/datasets/0x94t3z/fieldatlas-encyclopedia)). Encyclopedia and travel seeds come from Wikipedia's "vital articles" list in the repo, so questions are about well-known topics; first-aid seeds from a fixed list of first-aid topics.
2. **Question** written by the teacher model in one of several registers (casual phone typing, full sentence, two-part question, keyword search…), targeted at a label: fully answerable, partly answerable (one part not in the passage) or not answerable (same topic, detail not covered).
3. **Passages** retrieved by a replica of the app's own search (`tools/evaluation/e2e_test.py` from the repo: FTS5 ranking + planner keywords), mostly from the question's own collection plus 0–3 off-topic passages from another collection, with the seed passage at a random rank. This reproduces the noisy passage sets the model sees on the phone.
4. **Answer, label and key facts** written by the teacher under the benchmark's exact system prompt, with explicit rules: cite every sentence, copy numbers verbatim, no outside knowledge, exact refusal wording, gap sentence last for partial answers, no first-aid steps the passages do not give.
5. **Two filters.** (a) The benchmark's own scorer (`reader_bench.score_item`) plus style checks: every sentence cited, no citation to a missing passage, no number outside the passages, key facts present in cited passages, exact refusal phrasing. (b) An independent checker from a different model family reads the passages and the answer and fails anything unsupported, mislabelled, incomplete, unsafe or off-topic. About 70 % of generated answers pass both.
Teacher: `deepseek-flash` (DeepSeek API; its [terms](https://cdn.deepseek.com/policies/en-US/deepseek-open-platform-terms-of-service.html) allow using outputs to train other models). Checker: `openai/gpt-oss-120b` (Apache-2.0), used for pass/fail only; none of its text is in the data. No model whose terms forbid training on its outputs was used anywhere.
## Exclusions, so the benchmark stays clean
- No passage from any **dev-set article** (title or source URL) appears in any training item.
- No question is within a Jaccard similarity of 0.5 of any of the **~4,700 questions found anywhere under `docs/fieldtest/`** in the Field Atlas repo at commit `296b2fc` — the sealed test set is built from those field tests.
- The held-out set was split off once and frozen; later data top-ups could not move items between splits.
## Licences
- **Questions, answers, labels and key facts** (our annotations): CC BY-SA 4.0.
- **Passage texts** are reproduced from the Field Atlas knowledge packs and keep their original licences, with attribution in each pack's `ATTRIBUTION` file: Wikipedia and Wikivoyage text CC BY-SA 4.0; Ethereum EIPs/ERCs and consensus specs CC0; OpenStreetMap-derived text ODbL; biology passages are PubMed Central open-access articles under their individual licences (mostly CC BY, some CC BY-NC — check the `source` URL of an item before commercial reuse).
- Every item carries its passages' `source` URLs, so attribution is per item.
## Intended use and limits
Made for training and evaluating small grounded readers that must cite and abstain. English only. The reference answers are machine-written and machine-checked, not human-verified; the two-filter pass rate suggests they are accurate but not perfect. Key facts in `heldout.jsonl` were copied verbatim from the reference answers without alternative wordings, so they score stricter than the maintainer's dev set (see `fairscore.py` in the model repo for a wording-tolerant comparison).