MaxDevv/eval-Qwen3-235B-A22B-reasoning

这是一个重新格式化的评估结果数据集，针对Qwen-235B-A22B-Thinking模型，由Nous Hermes评估套件生成。数据集将原始评估输出（samples.jsonl）与对应的基准指标（metrics.json）合并为干净的Parquet表格，包含15个基准测试（如aime24、aime25、arenahard等），排除了creative-writing-v3和eqbench3。关键列包括完整的对话消息（messages）、用户查询（query）、模型响应（model_response）、真实答案（gold_answer，如果可用）、模型名称、温度等生成配置，以及每个基准的得分列（如math_pass@1:64_samples、eval_overall_winrate等）。数据集用于评估模型在数学、推理、知识问答等多个领域的性能，并提供了详细的得分和统计信息。

This dataset is a reformatted copy of eval results for Qwen-235B-A22B-Thinking, originally produced by the Nous Hermes evaluation suite. It combines raw evaluation outputs (samples.jsonl) with corresponding benchmark metrics (metrics.json) into clean, self-contained Parquet tables. It includes 15 benchmarks (e.g., aime24, aime25, arenahard, etc.), excluding creative-writing-v3 and eqbench3. Key columns include the full conversation messages, user query, model response, ground truth (when available), model name, temperature, and other generation configurations, along with per-benchmark score columns (e.g., math_pass@1:64_samples, eval_overall_winrate, etc.). The dataset is used to evaluate the models performance across various domains such as mathematics, reasoning, and knowledge Q&A, with detailed scores and statistics provided.