HealthCLBench / README.md
Tommy94's picture
Update README.md
da2a4e4 verified
|
Raw History Blame Contribute Delete
9.28 kB
metadata
pretty_name: HealthCLBench
language:
  - zh
license: cc-by-nc-sa-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: HealthCLBench.csv
task_categories:
  - question-answering
  - text-generation
size_categories:
  - 1K<n<10K
tags:
  - benchmark
  - healthcare
  - multi-turn-dialogue
  - context-learning
arxiv:
  - '2611.22362'

HealthCLBench

HealthCLBench evaluates whether a model can use earlier turns in a Chinese health conversation to answer the user's current question. Its 1,103 cases form 328 controlled clusters. Cases in a cluster share a setting, but a change in the evidence, the patient's state, or the user's goal can change what a good answer must address. Each case has its own medically reviewed scoring criteria. Paper

For example, three cases in cluster-0072 concern the same ointment. In case-0241, the user gives only its brand name. In case-0242, the user identifies terbinafine as the active ingredient and asks about applying it to cracked skin between the toes. In case-0243, the user asks about the same symptom but reports that the package lists methyl salicylate and menthol. The cluster tests whether the answer reflects what the user has established about the product in each conversation.

Run the evaluation

Dataset coverage

Release release-20261029 contains 1–5 cases per cluster, 1–50 user turns per case (median 4), and 4,931 rubric criteria. Each case has 2–14 criteria (median 4).

Medical scenario Label in the data Cases
Health assessment 健康评估 498
Clinical decision-making 诊疗决策 273
Health management 健康管理 164
Care triage 分诊就医 127
Medical explanation 医学解释 41

The primary context difficulty is labeled complex in 545 cases, fragmented in 305, conflicting in 81, and noisy in 65. These four families cover 14 subtypes, such as resolving an omitted reference, reconstructing a timeline, and reconciling conflicting facts. The remaining 107 cases carry the label 其他难点 (other difficulties).

Files and fields

Both files contain all 1,103 cases. Use the CSV for tabular analysis or the JSONL for direct access to the final question and cluster membership.

CSV column JSONL field Contents
num metadata.case_id Case identifier, e.g. CSV 241 corresponds to JSONL case-0241.
cluster_id metadata.cluster_id Cluster identifier, e.g. CSV 72 corresponds to JSONL cluster-0072.
messages messages Ordered conversation turns with role and content. JSONL stores content as lists of text blocks.
rubrics rubrics Case-specific criteria. Each has an id, a claim, a priority level (P0/P1/P2), and a type (positive/negative).
label metadata.label Medical scenario, primary and additional context difficulties and subtypes, departments, and diseases.

In the CSV, parse messages, rubrics, and label with json.loads. In the JSONL, metadata.query gives the verbatim final question and metadata.cluster_case_ids lists every member of the cluster. These fields make it possible to compare input conditions and select complete clusters. CSV download · JSONL download

Quick start

To inspect the three questions in the example cluster, install huggingface_hub with python -m pip install huggingface_hub, then run the following Python code. The revision pins the example to the release described above.

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="AQ-MedAI/HealthCLBench",
    repo_type="dataset",
    filename="HealthCLBench.jsonl",
    revision="fd3f2f4a75beb9d2e726c7904626374c98a06957",
)
with open(path, encoding="utf-8") as source:
    for line in source:
        case = json.loads(line)
        meta = case["metadata"]
        if meta["cluster_id"] == "cluster-0072":
            print(meta["case_id"], meta["query"])

Run the evaluation

Download the evaluation scripts and the release CSV into the same working directory. The script archive contains the project's public_eval runner, scoring code, and example configuration (code revision d7d3df2, Apache-2.0). Evaluation scripts (.zip) · Release CSV

Install and check the data

Run these commands from that directory using Python 3.11 or later.

unzip healthclbench-eval.zip
python3 -m venv .venv
. .venv/bin/activate
python -m pip install ./public_eval
hclbench validate --dataset HealthCLBench.csv
cp public_eval/config.example.json run.json

Validation should report 1,103 cases, 328 clusters, and 4,931 criteria. To try the scoring pipeline offline, run python public_eval/examples/demo.py --output results/demo. It uses two synthetic cases and fixed judge votes.

Evaluate a model and export scores

First, preview the example cluster with the supplied configuration. This step runs locally. With one evaluated model, the three cases in cluster-0072 require 6 Full/Query inference requests and up to 54 judge requests before retries.

hclbench-run --dataset HealthCLBench.csv --config run.json \
  --clusters cluster-0072 --output-dir results/sample --dry-run

For a live run, edit run.json. Under models, set the model you want to evaluate. Under judges, configure three distinct judges. For each entry, set model, the HTTPS Chat Completions base_url, its matching allowed_host, and the name of the credential environment variable in api_key_env. Set those variables in your shell or secret manager. The example uses HCLBENCH_API_KEY.

hclbench-run --dataset HealthCLBench.csv --config run.json \
  --clusters cluster-0072 --output-dir results/sample
hclbench-report --run-dir results/sample --output-dir results/sample-report

Read results/sample-report/README.md for the Full and Query case and cluster scores. The same directory contains the scores in summary.csv and the paired differences in the deltas field of summary.json. For the full benchmark, omit --clusters cluster-0072 and choose fresh run and report directories. To resume an interrupted evaluation, repeat its run command with --resume using the same dataset, configuration, and output directory.

Evaluation protocol

  1. Generate two answers per case. In Full, give the model the complete released conversation. In Query, give it only the text in metadata.query. Keep the model and generation settings fixed. Reserve rubrics for grading and labels for analysis.
  2. Grade the final answer. In both conditions, three judges see the full case evidence. Each votes independently on each criterion. Two matching Met or Not Met votes decide the outcome. A positive criterion passes on Met. For a negative criterion, Met means the prohibited defect occurred, so it passes on Not Met.
  3. Compute the case score. P0, P1, and P2 carry weights of 3, 2, and 1. A confirmed P0 failure gives the case a zero. Otherwise, subtract the weights of failed criteria from the weights of passed criteria, divide by total criterion weight, and floor at zero. Unresolved criteria contribute only to the denominator.
  4. Report case and cluster scores. The case metric averages case scores. For the cluster metric, take the geometric mean within each complete cluster, then average across clusters with equal cluster weight. A zero on any member gives that cluster a zero. Report both metrics on a 0–100 scale and compare Full with Query on the same cases.

Construction and intended use

The authors designed health scenarios, simulated conversations, and varied selected facts or task goals to construct related cases. A 70-person medical team annotated and revised the cases and their rubrics. Independent teams then reviewed a 25% quality-control sample and a 10% acceptance sample. The benchmark supports model comparison and analysis of context-use errors. Its simulated Chinese conversations and scenario counts reflect the benchmark design. They do not estimate clinical prevalence. Paper

License and citation

The dataset is released under CC BY-NC-SA 4.0. Attribute it to HealthCLBench, AQ-MedAI. License terms

@misc{yang2026healthclbenchevaluatingcontextlearninghealth,
  title={HealthCLBench: Evaluating Context Learning in Health Conversations through Controlled Case Clusters},
  author={Ting Yang and Zijing Hong and Chengjie Wang and Peng Wei and Chunxiao Guo and Yun Yue and Xurui Li and Jinqun Guan and Zixiao Chen and {Ant Group}},
  year={2026},
  eprint={2611.22362},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2611.22362}
}