TsinghuaC3I/ZEDA-Evaluation

该数据集包含60,000个来自数学、代码和聊天领域的提示，以及由教师MoE模型生成的相应自蒸馏展开。具体文件包括：ZEDA-prompts-60k.jsonl（用于展开和在策略蒸馏的源提示，来自AceReason-1.1-SFT和Llama-Nemotron-Post-Training-Dataset数据集）、ZEDA-Qwen3-30B-A3B-rollout-60k.jsonl（使用Qwen3-30B-A3B模型生成的展开结果，用于监督微调阶段）和ZEDA-GLM-4.7-Flash-rollout-60k.jsonl（使用GLM-4.7-Flash模型生成的展开结果，用于监督微调阶段）。The dataset consists of 60,000 prompts across math, code, and chat domains, along with their corresponding self-distillation rollouts generated by teacher MoE models. Files include: ZEDA-prompts-60k.jsonl (source prompts used for rollout and On-Policy Distillation, curated from AceReason-1.1-SFT and Llama-Nemotron-Post-Training-Dataset), ZEDA-Qwen3-30B-A3B-rollout-60k.jsonl (rollout results generated using the Qwen3-30B-A3B model, used for the Supervised Fine-Tuning stage), and ZEDA-GLM-4.7-Flash-rollout-60k.jsonl (rollout results generated using the GLM-4.7-Flash model, used for the SFT stage).