tanhaosheng/surgeon-tested-clinical-ai-benchmark

Surgeon-Tested Clinical AI Benchmark (TH-CAB v1.1)

An independent, reproducible evaluation of large language models (LLMs) on real, de-identified cancer cases — scored by a practicing surgeon item-by-item against current clinical guidelines.

• Homepage & full leaderboard: https://tanhaosheng.asia
• Methodology (citable authority, TH-CAB v1.1): https://tanhaosheng.asia/methodology/
• Open data layer: https://tanhaosheng.asia/data/

This is a benchmark / research dataset, not clinical validation, and not medical advice. All patient data is de-identified. AI is an assistant to clinicians, not a replacement.

What makes this dataset different

Most "AI beats doctor" headlines report a single run on a curated question. This dataset instead reports 20 repeated runs per model × case configuration, hand-scored by a surgeon on a fixed 16-item / 100-point rubric mapped to 2026 CBCS / CSCO 2024 / NCCN 2025 guidelines. The result is a distribution (mean ± SD), not a lucky single number — a clinical safety signal.

Every row is explicitly tiered by how the score was verified, so nobody has to take a number on faith:

Tier	Meaning
surgeon-verified	Surgeon-final-signed: the run was re-graded by Dr. Tan against the current rubric
staging-verified	Staging item verified on the current 16-item rubric; treatment/surgery items surgeon-final-signed
proxy	Keyword-proxy score — preliminary, may be revised after manual re-grading (see note)

Dataset at a glance

• 27 configuration rows (model × cancer)
• 11 cancer types: Breast, Cervical, Colorectal, Esophageal, Gastric, Liver, Lung, Ovarian, Pancreatic, Prostate, Thyroid
• Models across families: DeepSeek (V4 Flash, V4 Pro, R1, Quick, Web Fast), GLM (5.2, 5V-Turbo, 4-Flash), Gemini 3.1 Flash Lite, Agnes-2.0-flash, GPT-5, Zhipu / Zhipu-5.2
• ~574 scored runs total (sum of n; grows as new rows are published)

Schema

Each row of all-scores.json has the following fields:

Field	Type	Description
cancer	string	Cancer type (one de-identified real case per type)
model	string	Model name and interface (API / Web / agent wrapper)
mean	float	Mean score on the 16-item / 100-point rubric (n-run mean)
sd	float	Standard deviation across runs (null if not reported)
n	int	Number of repeated runs (typically 20)
stability	string	Stability class: Stable / Moderate / Unstable (null if not classified)
url	string	Relative link to the benchmark article with raw runs
runs	list	Per-run scores where published (empty if pending)
verified	bool	Whether the row has been human/LMJ-verified (vs raw proxy)
tier	string	Verification tier: surgeon-verified / staging-verified / proxy
note	string	Verification note / correction rationale (e.g. "Surgeon-final-signed correction, gold=pT1a")
benchmark	string	Rubric benchmark tag (e.g. current-rubric-surgeon-signed)
date	timestamp	Date of the verification pass (YYYY-MM-DD)

Corrections are disclosed, never silently merged. Where a proxy score was revised after manual re-grading, the old value and the reason are kept in note — see the correction policy on the site: https://tanhaosheng.asia/methodology/

Methodology in one line (TH-CAB v1.1)

1. One real, de-identified case per cancer type (chosen for guideline-relevant staging/treatment traps).
2. One locked prompt template, identical for every model — no model-specific tuning.
3. Fixed 16-item / 100-point rubric mapped to 2026 CBCS / CSCO 2024 / NCCN 2025.
4. 20 repeated runs per configuration to measure stability, not a single lucky answer.
5. Every run scored by hand by a surgeon (no LLM-as-judge, no auto-scoring) — proxy rows explicitly flagged.
6. Conflicts disclosed in visible correction boxes, never silently merged.

Cite the methodology page (https://tanhaosheng.asia/methodology/) when referencing any score.

Loading the data

from datasets import load_dataset

ds = load_dataset("tanhaosheng/surgeon-tested-clinical-ai-benchmark")
print(ds["train"][0])

Or read the JSON directly:

import json
rows = json.load(open("all-scores.json"))

Example analysis

import json
rows = json.load(open("all-scores.json"))

# Only surgeon-verified rows, best per cancer type
best = {}
for r in rows:
    if r.get("tier") == "proxy":
        continue
    c = r["cancer"]
    if c not in best or r["mean"] > best[c]["mean"]:
        best[c] = r
for c, r in sorted(best.items()):
    print(f"{c:12s} {r['model']:28s} {r['mean']:5.1f} [{r['tier']}]")

License

CC-BY-NC 4.0 — attribution required, non-commercial use. Please cite the dataset and link back to https://tanhaosheng.asia.

Citation

@misc{tan2026surgeontested,
  title        = {Surgeon-Tested Clinical AI Benchmark (TH-CAB v1.1)},
  author       = {Tan, Haosheng},
  year         = {2026},
  publisher    = {Surgeon-Tested AI},
  howpublished = {\url{https://tanhaosheng.asia}},
  note         = {De-identified real cancer cases, 16-item/100 rubric, repeated runs, surgeon-scored; tiered by verification}
}

Disclaimer

For educational and research purposes only, based on de-identified cases. Not individual medical advice. Consult a physician and follow current guidelines (NCCN / CSCO / CBCS). Nothing on this dataset or the associated site is a substitute for professional medical judgment.