latent-sft-eval-benchmarks
收藏资源简介:
Latent-SFT评估基准是一个用于评估Latent-SFT轨迹生成和模型诊断的标准化数据集集合。该数据集整合了多个学术评估基准,包括MMLU-Pro、GPQA、SuperGPQA和BBEH,旨在支持批量思维链(CoT)轨迹生成任务。数据以JSONL格式提供,分为主规范化文件和审计文件:主文件包含标准化的问题(problem)和答案(answer)字段,适用于模型推理;审计文件则保留原始数据源的ID、类别、领域等元数据,用于追溯和分析。数据集规模具体为:MMLU-Pro包含验证集70个样本和测试集12032个样本;GPQA钻石子集包含198个样本;SuperGPQA包含26529个样本;BBEH包含4520个样本。使用要求:在生成CoT轨迹时,最终框定答案应仅出现在响应或cot_answer字段中,推理部分不得包含额外答案。该数据集为纯文本格式,排除了依赖图像的TheoremQA样本,适用于文本生成和问答任务的研究与评估。
The Latent-SFT Evaluation Benchmark is a standardized dataset collection for evaluating Latent-SFT trajectory generation and model diagnosis. It integrates multiple academic evaluation benchmarks, including MMLU-Pro, GPQA, SuperGPQA, and BBEH, aiming to support batch Chain-of-Thought (CoT) trajectory generation tasks. The data is provided in JSONL format, divided into main normalized files and audit files: the main files contain standardized problem and answer fields suitable for model inference; the audit files retain metadata such as original data source IDs, categories, and domains for traceability and analysis. The dataset scale is as follows: MMLU-Pro includes a validation set of 70 samples and a test set of 12,032 samples; the GPQA diamond subset contains 198 samples; SuperGPQA contains 26,529 samples; and BBEH contains 4,520 samples. Usage requirements: when generating CoT trajectories, the final framed answer should only appear in the response or cot_answer field, and the reasoning part must not contain additional answers. The dataset is in plain text format, excluding image-dependent TheoremQA samples, and is suitable for research and evaluation in text generation and question-answering tasks.
数据集概述
数据集名称:Latent-SFT Evaluation Benchmarks
许可证:未知
语言:英语(en)
任务类别:问答、文本生成
标签:mmlu-pro, gpqa, supergpqa, bbeh
数据集描述
该数据集是用于Latent-SFT轨迹生成和模型诊断的评估数据集,经过处理打包为批量CoT(思维链)轨迹生成格式。提示要求仅在生成响应或cot_answer中包含最终框出的答案,推理部分不应包含多余的框出答案。
文件结构
| 文件 | 行数 | 键值 |
|---|---|---|
| mmlu_pro_validation.jsonl | 70 | answer, problem |
| mmlu_pro_validation_audit.jsonl | 70 | answer, answer_index, category, question_id, src |
| mmlu_pro_test.jsonl | 12032 | answer, problem |
| mmlu_pro_test_audit.jsonl | 12032 | answer, answer_index, category, question_id, src |
| gpqa_diamond.jsonl | 198 | Correct Answer, Explanation, High-level domain, Incorrect Answer 1, Incorrect Answer 2, Incorrect Answer 3, Question, Record ID, Subdomain |
| gpqa_diamond_unshuffled_normalized.jsonl | 198 | answer, problem |
| supergpqa_all.jsonl | 26529 | answer, problem |
| supergpqa_all_audit.jsonl | 26529 | answer, answer_text, difficulty, discipline, field, subfield, uuid |
| bbeh_all.jsonl | 4520 | answer, problem |
| bbeh_all_audit.jsonl | 4520 | index, mini, target, task |
注意事项
- 主要的标准化文件使用
problem和answer键(如可用)。 *_audit.jsonl文件保留了原始ID和数据集特定元数据。gpqa_diamond.jsonl采用GPQA原始格式列;gpqa_diamond_unshuffled_normalized.jsonl为标准化多项选择视图。- 故意排除TheoremQA,因为此纯文本包不包含其图像依赖样本。
- 本包不包含任何生成的CoT轨迹。
数据格式
所有JSONL文件以.jsonl.gz压缩格式上传,可通过Hugging Face datasets库直接流式读取。生成的最终框出答案应仅出现在response/cot_answer中,不应出现在推理部分。




