LogiMed-RoB Dataset

LogiMed-RoB accompanies the paper “Can LLMs Follow Medical Expert Logic? A Benchmark for Hierarchical Logical Consistency in Risk-of-Bias Assessment.” It evaluates whether large language models follow the hierarchical expert logic used in clinical Risk-of-Bias (RoB) assessment, rather than merely matching final labels.

The release contains 860 randomized controlled trials (RCTs) and 14,820 structured queries across two complementary tracks. Gold-standard labels originate from expert assessments in Cochrane systematic reviews, peer-reviewed publications, ROBIN, and RoBBR. No AI-generated pseudo-labels are used.

Dataset configurations

Configuration	Standard	RCTs	Items	Structured queries	Sources
track_a	Cochrane RoB 2.0	201	626	13,772	Cochrane (467), Figshare (159)
track_b	Cochrane RoB 1.0	659	1,048	1,048	ROBIN (713), RoBBR (335)

Track A contains 22 rule-based signaling queries per assessment. Track B contains one evidential-faithfulness query per item.

Files

File	Description
RCT_ROB2_final_dataset.json	Track A: RoB 2.0 outcome-level assessments with full RCT text, outcome descriptions, domain judgments, expert rationales, and overall risk labels.
ssr_dataset_merged.json	Track B: RoB 1.0 evaluation items with signaling questions, expert labels, evidence windows, answers, and rationales.

Both files are UTF-8 JSON arrays and are exposed as a test split because LogiMed-RoB is an evaluation benchmark rather than a training corpus.

Loading the dataset

from datasets import load_dataset

track_a = load_dataset(
    "BUPT-Reasoning-Lab/LogiMed-RoB",
    "track_a",
    split="test",
)

track_b = load_dataset(
    "BUPT-Reasoning-Lab/LogiMed-RoB",
    "track_b",
    split="test",
)

Download the complete repository with the Hugging Face CLI:

hf download BUPT-Reasoning-Lab/LogiMed-RoB \
  --repo-type dataset \
  --local-dir ./LogiMed-RoB

Data format

Track A: RCT_ROB2_final_dataset.json

Each entry contains the following fields:

{
  "id": "<uuid>",
  "study": "<study identifier>",
  "full_paper": "<Markdown text of the RCT paper>",
  "outcome": "<target outcome>",
  "randomisation_process_judgment": "<risk label>",
  "randomisation_process_reasoning": "<expert rationale>",
  "intended_interventions_judgment": "<risk label>",
  "intended_interventions_reasoning": "<expert rationale>",
  "missing_outcome_data_judgment": "<risk label>",
  "missing_outcome_data_reasoning": "<expert rationale>",
  "measurement_outcome_judgment": "<risk label>",
  "measurement_outcome_reasoning": "<expert rationale>",
  "selection_reported_result_judgment": "<risk label>",
  "selection_reported_result_reasoning": "<expert rationale>",
  "overall_risk": "<risk label>"
}

The RoB 2.0 domains are:

1. Bias arising from the randomisation process.
2. Bias due to deviations from intended interventions.
3. Bias due to missing outcome data.
4. Bias in measurement of the outcome.
5. Bias in selection of the reported result.

Track B: ssr_dataset_merged.json

Each entry contains the following fields:

{
  "id": "<uuid>",
  "study": "<study identifier>",
  "bias": "<bias domain>",
  "question": "<signaling question>",
  "label": "low | high | unclear",
  "context_list": ["<evidence sentence>", "..."],
  "answers": [
    {
      "answer": "yes | no | unclear",
      "reason": "<expert rationale>"
    }
  ]
}

Quality assurance

• Track A: all 626 items were manually extracted and calibrated against published expert conclusions; 467 are allocated to Cochrane and 159 to Figshare.
• Track B: approximately 20% of the release (210 items) was independently re-verified by medical annotators, with 93.5% agreement on evidence-window accuracy.
• No AI-generated pseudo-labels are included in either track.

Responsible use and source terms

LogiMed-RoB is intended for research on model evaluation, evidence-based medicine, and clinical reasoning. It must not be used as an autonomous clinical decision-making system.

The release incorporates material derived from Cochrane Library, Figshare, ROBIN, and RoBBR. Use and redistribution of individual records remain subject to the terms of their original sources; therefore the dataset card uses the Hugging Face other license designation. Consult the source projects and accompanying paper for provenance details.

Project resources

• Code and documentation: https://github.com/BUPT-Reasoning-Lab/LogiMed-RoB
• Dataset: https://huggingface.co/datasets/BUPT-Reasoning-Lab/LogiMed-RoB

Citation

@misc{huang2026logimedrob,
  title  = {Can LLMs Follow Medical Expert Logic? A Benchmark for Hierarchical Logical Consistency in Risk-of-Bias Assessment},
  author = {Huang, Jiayu and Tang, Zichen and Ling, Qianhui and Kuang, Zemin and E, Haihong},
  year   = {2026},
  url    = {https://github.com/BUPT-Reasoning-Lab/LogiMed-RoB}
}