delphi-1e23-25b-stageE-rl-eval-artifacts
收藏资源简介:
delphi-1e23 (25B) Stage-E RL 评估数据集是一个用于评估大型语言模型数学推理能力的数据集。它包含对一系列250亿参数语言模型(包括一个监督微调基线模型和四个不同强化学习目标微调的模型)在两个数学推理基准任务(MATH500和gsm8k)上进行性能评估的原始输出结果。数据集由lm-eval框架生成,提供两种文件类型:一是包含任务整体聚合指标(如准确率)和评估配置的JSON文件;二是记录每个评估样本详细信息(如原始问题、标准答案、模型输出、提取答案和正确性判断)的JSONL文件。评估设置具体为:MATH500任务采用0-shot设置,最大生成令牌数3072;gsm8k任务采用5-shot设置,最大生成令牌数2048。所有评估通过vLLM服务进行。该数据集旨在支持对模型数学问题求解能力的细粒度分析和比较,适用于研究社区进行模型评估、基准测试和强化学习微调策略影响分析。
The delphi-1e23 (25B) Stage-E RL evaluation dataset is designed for assessing the mathematical reasoning capabilities of large language models. It contains raw output results from performance evaluations of a series of 25-billion-parameter language models (including one supervised fine-tuned baseline model and four models fine-tuned with different reinforcement learning objectives) on two mathematical reasoning benchmark tasks (MATH500 and gsm8k). Generated using the lm-eval framework, the dataset provides two types of files: JSON files with overall aggregated metrics (e.g., accuracy) and evaluation configurations, and JSONL files with detailed information for each evaluation sample (such as original questions, reference answers, model-generated outputs, extracted answers, and correctness judgments). The evaluation settings are specified as: MATH500 task uses 0-shot setup with a maximum token generation of 3072, and gsm8k task uses 5-shot setup with a maximum token generation of 2048. All evaluations are conducted via a configured vLLM service. This dataset aims to support fine-grained analysis and comparison of model performance in mathematical problem-solving, and is suitable for the research community in model evaluation, benchmarking, and analyzing the impact of reinforcement learning fine-tuning strategies.
数据集概述
数据集名称: laion/delphi-1e23-25b-stageE-rl-eval-artifacts
许可证: Apache-2.0
标签: evalchemy, delphi, rl, math-eval
描述: 该数据集包含 delphi-1e23 25B 模型在 Stage-E RL(强化学习)扫描中的原始 evalchemy 评估结果(基于 lm-eval v0.4.12)。覆盖了5个模型(1个SFT基线和4个RL变体)在2个数学任务上的评估输出。
模型
- 基线模型:
laion/delphi-1e23-wc50m-warmup-levanter-sft - RL变体:
laion/delphi-1e23-wc50m-rl-d1-rlvrmathlaion/delphi-1e23-wc50m-rl-d2-ifevallaion/delphi-1e23-wc50m-rl-d3-dapomathlaion/delphi-1e23-wc50m-rl-d4-math500-32gpu
评估任务
- MATH500: 0-shot评估,最大生成长度
EVAL_MAX_GEN_TOKS=3072。 - gsm8k: 5-shot评估,最大生成长度
EVAL_MAX_GEN_TOKS=2048(避免4k上下文溢出问题)。
文件结构
每个模型和任务组合生成两种文件:
-
<TASK>_<MODEL>_results.json: 包含聚合指标和完整运行配置。指标包括:- MATH500: 准确率(accuracy)
- gsm8k: 精确匹配(exact_match, 包括flexible和strict两种模式)
-
<TASK>_<MODEL>_samples.jsonl: 逐样本记录,包含:- 问题(problem)
- 标准答案(gold)
- 模型输出(model_output)
- 提取的答案(extracted answer)
- 正确性标记(correctness)
评估配置
- 推理方式: vLLM,张量并行度(TP)为6(delphi模型有42个注意力头,TP必须整除42)。
- 对话模板: delphi_v0(
skip_special_tokens=false) - 最大模型长度:
max_model_len=4096




