refute
收藏资源简介:
REFUTE(证据推理)是一个基准数据集,旨在评估语言模型的科学推理能力和认知诚实性,核心在于区分模型的‘技能’(如撰写批判性文本)与‘诚实性’(如对证据强度的正确校准),即模型在证据薄弱时是否过度自信。数据集基于2025-2026年发表的真实科学论文构建,包含两个主要版本:v2版本(2026年发布)包含240个自动评分的多项选择题,分为四个任务类型:1) 知识回忆:60个闭卷四选一题目,测试对近期研究发现的记忆;2) 证伪选择:60个题目,要求选择能具体证伪某个主张的观察结果;3) 标题校准:40个题目,区分谨慎结论与 subtly hyped 的结论;4) 缺陷识别:80个题目,在缺陷 subtle 的情况下选择最合理的摘要。v1版本则侧重于开放式文本生成,包含60-120个需要对真实论文进行批判性写作的任务,以及74个带有客观‘合理/有缺陷’标签的小故事,用于校准评估。数据集适用于文本生成、问答和文本分类任务,特别服务于科学推理、认知论、评估、校准和证伪等研究领域。数据以纯文本形式提供,总规模小于1000个样本。
REFUTE (Evidence Reasoning) is a benchmark dataset designed to evaluate language models scientific reasoning abilities and cognitive honesty. The core idea is to distinguish between a models skill (such as writing critical texts) and honesty (such as correctly calibrating evidence strength), i.e., whether the model is overconfident when evidence is weak. The dataset is constructed based on real scientific papers published in 2025-2026 and includes two main versions. The v2 version (released in 2026) contains 240 automatically scored multiple-choice questions, divided into four task types: 1) Knowledge recall: 60 closed-book four-option questions testing memory of recent research findings; 2) Falsification selection: 60 questions requiring selection of observations that specifically falsify a claim; 3) Title calibration: 40 questions distinguishing cautious conclusions from subtly hyped ones; 4) Flaw identification: 80 questions selecting the most reasonable summary when flaws are subtle. The v1 version focuses on open-ended text generation, including 60-120 tasks requiring critical writing on real papers, and 74 short stories with objective reasonable/flawed labels for calibration evaluation. The dataset is suitable for text generation, question answering, and text classification tasks, particularly serving research areas such as scientific reasoning, epistemology, evaluation, calibration, and falsification. Data is provided in plain text format, with a total size of less than 1000 samples.
数据集概述:REFUTE(Reasoning Over Evidence)
- 名称:REFUTE(Reasoning Over Evidence)
- 语言:英语(en)
- 许可证:Apache-2.0
- 任务类别:文本生成、问答、文本分类
- 标签:基准测试、科学推理、认知论、评估、校准、证伪、纯文本
- 大小:n<1K(少于1000条样本)
- 发布机构:BGPT(https://bgpt.pro/)
核心目标
REFUTE基准测试旨在评估前沿语言模型对2025–2026年真实论文的科学判断能力,重点考察模型是否“诚实”(即对自身认知的准确性),而非仅仅考察其能否给出正确答案。它区分“批判技能”和“认知诚实性”两个维度。
数据集构成
REFUTE包含两个版本(v1和v2),共7个配置,均为训练集(split: train)。
v2(2026年版本)
- 风格:多项选择题,可自动评分,无需人工判断。
- 总量:240道题目,涵盖4种问题类型:
| 问题类型 | 配置名称 | 题量 | 说明 |
|---|---|---|---|
| 你知道研究发现吗? | refute_knowledge | 60 | 考察对2025–2026年最新研究结果的回忆(封闭式4选1,精确匹配) |
| 什么能证明它错了? | refute_falsifier_choice | 60 | 从近似干扰项中选出具体的证伪观察 |
| 标题是否公允? | refute_overclaim_choice | 40 | 区分谨慎结论与过度营销 |
| 你能识别糟糕的科学吗? | refute_discrimination_hard | 80 | 在细微缺陷中选出最合理的总结 |
- 特征字段:id, task, prompt, answer, options, flaw_type, rubric_type
- 现有前沿模型平均正确率:76%;其中“识别糟糕科学”最困难(67%)。
v1(原始版本)
- 风格:开放式生成式批评,需人工或标准评分。
- 总量:3个配置,用于评估生成式科学写作与校准。
| 配置名称 | 题量 | 说明 |
|---|---|---|
| refute_120 | 120 | 五大任务类型(每种24条),纯文本生成式批评 |
| refute_hard_60 | 60 | 挑战性子集,用于生成式排名 |
| refute_soundness | 74 | 包含客观的“合理/有缺陷”标签,用于校准评估 |
- 特征字段(refute_120 / refute_hard_60):id, task, paper_title, paper_date, input, reference, rubric_type, max_score, rubric, scoring_notes, source_hash
- 特征字段(refute_soundness):id, label, flaw_type, summary, source, flaw_tell
数据文件格式与加载示例
所有数据文件为JSONL格式,可通过Hugging Face datasets 库加载。
-
加载v2(以
refute_knowledge为例): python from datasets import load_dataset items = load_dataset("BGPT-OFFICIAL/refute", "refute_knowledge", split="train") -
加载v1(以
refute_hard_60为例): python hard = load_dataset("BGPT-OFFICIAL/refute", "refute_hard_60", split="train")
排行榜与评分
- Truth Score:综合知识、不确定性诚实度、缺陷识别和批判技能的0–100评分。
- 当前领先模型:Claude-Opus-4.7(73分),其次为Gemini-3.1-Pro(70分)、Claude-Opus-4.6(68分)。
- 完整排行榜:https://huggingface.co/spaces/BGPT-OFFICIAL/refute-leaderboard
相关资源
- 技术报告:TECHNICAL_REPORT.md
- 评估协议:metadata/eval_protocol_mcq_v2.json
- 集成指南:INTEGRATORS.md
- FAQ:FAQ.md
- 引用: bibtex @misc{bgpt_refute_v2_2026, title = {REFUTE: Reasoning Over Evidence Benchmark}, author = {{BGPT Team}}, year = {2026}, url = {https://huggingface.co/datasets/BGPT-OFFICIAL/refute} }





