--- pretty_name: MedPRMBench Test language: - en task_categories: - question-answering tags: - medical - reasoning - process-reward-model configs: - config_name: default data_files: - split: test path: data/test.parquet --- # MedPRMBench test set MedPRMBench is a step-level benchmark for checking clinical reasoning chains. This repository contains the **test split only**: 6,500 questions, each with one original reasoning chain and one corrupted chain. Labels identify error steps and clinical error types. The data comes from the ICLR 2027 submission archive retained on 2026-09-26 and assembled on 2026-10-09. The source archive marks this split `RELEASE_CANDIDATE_NOT_YET_AUTHORITATIVE`. Its status and source checksum are recorded in `manifest.json`. No blanket license for the seven upstream question sources is asserted here; maintainers should confirm their redistribution terms before a public release. ## Load ```python from datasets import load_dataset test = load_dataset("YOUR_ORG/MedPRMBench-Test", split="test") print(len(test)) # 6500 ``` Replace `YOUR_ORG` with the Hugging Face account used for upload. The dataset viewer reads `data/test.parquet`. The untouched archived JSONL is available at `raw/test_benchmark.jsonl`; its SHA-256 is in `manifest.json`. ## Record structure | Field | Meaning | | --- | --- | | `id`, `dataset`, `dataset_type`, `split` | Record ID, source dataset, Class A (expert-authored) or B (generated), and test split. | | `question`, `options`, `correct_answer` | Question, answer options, and reference answer. | | `original_steps`, `corrupted_steps`, `corrupted_answer` | Paired reasoning chains and the corrupted-chain answer. | | `error_steps`, `error_step_code_map` | Gold error-step indices and their type mapping; indices are **1-based**. | | `error_codes`, `error_categories` | Error types and their broad categories. | | `modified_steps`, `error_injection_reason` | Construction metadata. | | `severity_level`, `severity_score` | Rule-defined construction severity; not a physician-rated patient-harm estimate. | | `answer_changed` | Whether the answer changed. In Parquet, 37 source values of `"unknown"` are null. | | `vote_error_failed`, `vote_reasoning_adopt_count`, `vote_reasoning_reject_count` | Optional review metadata. | The original JSONL has 47 dictionary-form `original_steps` elements across six records. In Parquet, their text is extracted so every reasoning step is a string. The original values remain in the raw JSONL. Parquet represents absent optional fields as null; other field content is unchanged. There are 3,711 Class A and 2,789 Class B records. ## Error codes | Code | Type | Code | Type | | --- | --- | --- | --- | | S-1 | Non-Redundancy | S-2 | Non-Circular Logic | | R-1 | Evidence-Based Soundness | R-2 | Step Consistency | | R-3 | Contextual Applicability | R-4 | Confidence Invariance | | R-5 | Safety Awareness | R-6 | Information Grounding Compliance | | R-7 | Trajectory Reasoning | E-1 | Prerequisite Sensitivity | | E-2 | Deception Resistance | E-3 | Multi-Solution Consistency | | E-4 | Quantitative Correctness | E-5 | Differential Diagnosis Coverage | Use the held-out split for evaluation. Selecting prompts, thresholds, or checkpoints on this split would compromise that evaluation.