Douxgen/OpenVetQA-sft-v1

OpenVetQA v1.0

A veterinary multiple-choice SFT dataset where every item is traceable to a CC-BY source article. Questions were generated from — and verified against — single passages from PMC open-access reviews and guidelines, then filtered by structural gates and blind cross-family LLM judging.

2,651 items · CC BY-NC 4.0 · Trust report · Gemma adapter · Qwen adapter

Results

Two 4B models were fine-tuned on the train split and evaluated on two held-out tiers.

Eval tier	Gemma-3-4B base → tuned	Qwen3-4B base → tuned
dev2 (157 items)	64.97% → 75.80% (+10.83)	66.88% → 82.17% (+15.29)
VetQA-Silver (146 items, holdout)	69.18% → 77.40% (+8.22)	76.71% → 84.93% (+8.22)

VetQA-Silver documents were carved out by sha256(source_doc_uid) % 20 == 0 before any generation ran. No generator, judge, or training run ever saw them, so the +8.2 gain on both families measures generalization rather than memorization.

Adapters: Douxgen/OpenVetQA-gemma3-4b-v1, Douxgen/OpenVetQA-qwen3-4b-v1.

Data

Split	Items
train	2,494
dev	157

2,247 items are synthetic (7 generator models); 404 are earlier draft items that were re-verified from scratch and merged in.

Each row is a chat-format conversation with embedded provenance:

{
  "messages": [
    {"role": "system",
     "content": "You are assisting a veterinary professional..."},
    {"role": "user",
     "content": "When performing hot-iron disbudding in calves...\nA. ...\nB. ..."},
    {"role": "assistant",
     "content": "A. Combining a local anesthetic with NSAIDs\nRationale: ..."}
  ],
  "provenance": {
    "chunk_id": "pmc:7dda00c8ba179f13e629",
    "source_doc_uid": "pmc:PMC7095864",
    "license": "cc-by",
    "citation": "J Vet Intern Med. 2020; 34(3):1062-1077"
  }
}

The user turn holds the question and options; the assistant turn holds the answer and a short rationale grounded in the source passage. The dev split is separated by sha256(source_doc_uid) % 100 >= 95, so no document appears in both splits.

How the data was made

1. 1,700 passages sampled deterministically from 114,806 eligible chunks of a frozen 946,884-chunk PMC store (reviews, guidelines, systematic reviews; 2015+; seven species groups).
2. Seven LLM generators wrote strict-JSON MCQs from their passage: 3,406 raw items.
3. Mechanical gates — option count and uniqueness, key validity, rationale minimums, control-character ban — plus 8-gram decontamination against each item's own source and the Merck Veterinary Manual, plus global dedupe. 1,077 items rejected.
4. Blind cross-family judging — a model from a different family than the generator scored every item without the answer key, on grounding, domain, and single-answerability. 88% inter-judge agreement across 1,427 items double-judged.
5. 404 legacy items re-verified through the same gauntlet and merged.

Same-family judging is banned in code. In a controlled bakeoff, blind judges caught 5/5 planted fabrications that the mechanical gates missed. A 20-item human audit (15/20 pass) produced fixes that are now permanent pipeline rules.

Eval tiers shipped alongside

Tier	Items	Construction
RegulatoryGold	3,385	regex over APVMA public registration records; no LLM involvement
VetQA-Silver	200+	generated and judged from holdout documents excluded from SFT
ExternalEval	—	deferred; author-constructed fallback documented in the trust report

Usage

from datasets import load_dataset

ds = load_dataset("Douxgen/OpenVetQA-sft-v1")
print(ds["train"][0]["messages"])

Limitations

• Items are grounding-verified against their sources, not veterinarian-validated. Clinical truth is inherited from the peer-reviewed sources.
• Some items are simple label lookups; the eval tiers exist to measure reasoning beyond that.
• The 404 legacy items use a 5-option format; synthetic items use 4.
• Sources are 2015+ publications. Verify doses against current labels.
• English only. Single training seed.

License

CC BY-NC 4.0. Sources are CC-BY PMC open-access articles; the per-item license and citation are embedded in every row.

Citation

@dataset{openvetqa2026,
  title  = {OpenVetQA v1.0: a provenance-complete veterinary MCQ dataset},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Douxgen/OpenVetQA-sft-v1}
}