bowyerlab/clinical-panel-corpus

Clinical Panel Corpus

A corpus of independent, round-one (pre-discussion) multi-role clinical panel responses: clinical questions in paired closed/open form, answered by a five-role panel of LLM agents, sampled 30x per case across multiple models.

v0.1 preview. This release contains the round-one (independent, pre-discussion) panel responses only. The panel design is being revised; a later version will add deliberation rounds and abstention labels and will supersede this one. The generation and analysis code is not yet public. Expect the layout to stay stable, but treat v0.1 as a preview rather than the version of record.

⚠️ AI-generated outputs — not clinical advice. Every response was produced by a language model for research on panel behaviour. None of it is validated for, or intended for, clinical use.

At a glance

Cases	300 closed (MedXpertQA-R, MedXpertQA-U MCQ); 268 also in open (de-optioned) form
Roles	Diagnostician, Investigator, Therapeutist, Safety and Pharmacology specialist, Mechanism specialist (5)
Sampling	30 independent realisations per case; temperature 0.7; root seed 20260428
Models	claude-haiku-4-5-20251001, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-3.5-flash
Total	340,800 panel responses (4 models × 2 arms × 5 roles × 30 runs)
Per response	primary_recommendation, key_finding, contraindication_flag, reasoning
Files	2,312 (embeddings: 37, loaded: 3, responses: 2272) + RELEASE_MANIFEST.json

Layout

Three top-level folders + RELEASE_MANIFEST.json (sha256 of every file):

• responses/ — one JSONL trace per case at responses/<model_tag>/panel_runs/headline__<case_id>__<spec-hash>.jsonl; <model_tag> fuses model and arm (closed vs -open). Because <case_id> contains a / (e.g. MedXpertQA-R/Text-702), each trace nests one directory deeper. Each trace is a header row followed by one response row per (role × realisation). A validity flag is computed at load time, not stored.
• embeddings/ — optional float32 .npz grids over the question and the response fields (text-embedding-005).
• loaded/ — rehydrated case content (the closed + de-optioned open questions), as cases.parquet and cases_open_ended.parquet.

Loading

Each trace mixes a header row with response rows, so the generic datasets auto-loader is not the intended entry point. Download the repo and read the traces directly:

import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download("bowyerlab/clinical-panel-corpus", repo_type="dataset",
                              allow_patterns=["responses/gemini-2_5-flash-range/**"]))

for trace in root.glob("responses/*/panel_runs/**/*.jsonl"):
    rows = [json.loads(line) for line in trace.open()]
    header, responses = rows[0], rows[1:]
    for r in responses:
        # r["agent_id"], r["run_idx"], r["seed"], r["model_version"], token counts, timestamp
        try:
            fields = json.loads(r["text"])   # the four structured fields
        except json.JSONDecodeError:
            fields = None                    # refusal / malformed output: kept as observed

Header rows carry case_id, agent_ids, n_runs, panel_spec_name and panel_spec_hash. Response rows carry agent_id, agent_idx, run_idx, seed, model_version, resolved_model_version, input_tokens, output_tokens, thinking_tokens, timestamp and text (the model's raw output, normally a JSON object with the four fields above). Join to the question content on case_id via loaded/cases.parquet (closed arm) or loaded/cases_open_ended.parquet (open arm).

Scope — not a correctness benchmark

Each role gives its own domain assessment (a diagnosis, the next investigation, a management step, a safety flag, a mechanism), not the source MCQ's answer option — so the gold answer is not a correctness label for these responses, by construction. Pair with an external reference standard if you need correctness. What is validated is the data's integrity and behaviour (coverage, schema validity, refusal/degeneracy rates, role fidelity, diversity, stochastic variation).

Licensing

• Responses, corpus structure, de-optioned rewrites: CC-BY-4.0.
• Source questions: MedXpertQA-R, MedXpertQA-U via MedAgentsBench (dataset commit 87e99bf4), MIT-licensed and redistributed under MIT.

Citation

No DOI or paper yet. Until there is one, please cite the dataset directly:

Bowyer, S. (2026). Clinical Panel Corpus (v0.1, preview) [Dataset]. Imperial College London. https://huggingface.co/datasets/bowyerlab/clinical-panel-corpus

@misc{bowyer2026clinicalpanelcorpus,
  author       = {Bowyer, Stuart},
  title        = {Clinical Panel Corpus},
  year         = {2026},
  version      = {0.1},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/bowyerlab/clinical-panel-corpus}},
  note         = {Preview release}
}

Ethics

Source items are exam/benchmark questions — no real patient data, no PII. All responses are model-generated and labelled as not clinical advice.