SCOPE
收藏资源简介:
SCOPE数据集由南洋理工大学、Shopee和SEA Group的研究人员创建,包含196,000个数学推理样本,每个样本都有1.4M个步骤级别的标签。该数据集通过SCOPE方法构建,该方法将自然语言推理步骤转换为代码并标准化,然后合并等效步骤以构建前缀树。SCOPE数据集旨在解决现有过程标注方法计算成本高昂的问题,通过引入基于压缩的方法,将计算复杂性从O(NMK)降低到O(N)。SCOPE数据集已证明,在Best-of-N策略和ProcessBench 1等评估设置下,基于该数据集训练的PRM始终优于现有的自动标注方法。
The SCOPE dataset was created by researchers from Nanyang Technological University, Shopee, and SEA Group. It contains 196,000 mathematical reasoning samples, with 1.4 million step-level labels per sample. The dataset is constructed via the SCOPE method, which converts natural language reasoning steps into code, standardizes these steps, and merges equivalent ones to build a prefix tree. The SCOPE dataset aims to address the high computational cost issue of existing process annotation methods, introducing a compression-based approach that reduces the computational complexity from O(NMK) to O(N). It has been demonstrated that PRMs trained on this dataset consistently outperform existing automatic annotation methods under evaluation settings such as the Best-of-N strategy and ProcessBench 1.
SCOPE数据集概述
环境配置
- Python版本:3.10
- 依赖库:
- torch
- flash-attn
- vllm
- requirements.txt中列出的其他依赖
模型训练流程
- 生成响应
- 执行脚本:
generate_response.py
- 执行脚本:
- 生成代码
- 执行脚本:
generate_code.py
- 执行脚本:
- 代码压缩
- 执行脚本:
code_process.py - 参数:
--code_path:代码路径--use_code:使用代码--use_ast:使用AST
- 执行脚本:
- 训练模型
- 执行命令:
accelerate launch --config_file configs/accelerate_config.yaml train_reward.py
- 执行命令:
评估
BON评估
- 为每个响应生成步骤奖励
- 执行脚本:
gen_reward.py - 参数:
--model_path:模型路径--tokenizer_path:Qwen/Qwen2.5-Math-7B-Instruct
- 执行脚本:
- 评估BON
- 执行脚本:
prm_maj_eval.py - 参数:
--model_name:模型名称
- 执行脚本:
ProcessBench评估
- 执行脚本:
run_eval_our.py - 参数:
--model_path:模型路径--tokenizer_path:Qwen/Qwen2.5-Math-7B-Instruct--output_dir:输出目录

- 1SCOPE: Compress Mathematical Reasoning Steps for Efficient Automated Process Annotation南洋理工大学, 新加坡; Shopee Pte. Ltd, 新加坡; SEA Group, 新加坡 · 2025年



