marin-community/open-thoughts-4-30k-math-qwen3-235b-a22b-annotated

该数据集是Qwen3-235B-A22B注释版本的一个子集，包含29,963个样本，来源于[marin-community/open-thoughts-4-math-qwen3-32b-annotated](https://huggingface.co/datasets/marin-community/open-thoughts-4-math-qwen3-32b-annotated)，最初由OpenThoughts4团队从[mlfoundations-dev/hero_run_4_math](https://huggingface.co/datasets/mlfoundations-dev/hero_run_4_math)整理而来。数据集提供了[Qwen/Qwen3-235B-A22B-FP8](https://huggingface.co/Qwen/Qwen3-235B-A22B-FP8)的响应，生成参数为temperature = 0.8和max output tokens = 16000。数据集的结构包括多个列，如instruction_seed（原始数学问题/问题文本）、_source（原始数据集来源）、gpt41_mini_response（GPT-4.1 Mini生成的参考解决方案）、__original_row_idx（原始源数据集的行索引）、length（响应的令牌计数）、ms_id（唯一样本标识符）、generated_text（父数据集中由Qwen3-32B生成的响应，包含带有<think>标签的思维链）、qwen235b_generated_text（由Qwen3-235B-A22B-FP8生成的响应，包含带有<think>标签的思维链）和conversations（聊天格式的提示和响应）。与父数据集的主要区别在于添加了qwen235b_generated_text列，该列包含由Qwen3-235B-A22B-FP8生成的响应。

This dataset is the Qwen3-235B-A22B annotated version of a 29,963 sample subset from [marin-community/open-thoughts-4-math-qwen3-32b-annotated](https://huggingface.co/datasets/marin-community/open-thoughts-4-math-qwen3-32b-annotated), originally derived from [mlfoundations-dev/hero_run_4_math](https://huggingface.co/datasets/mlfoundations-dev/hero_run_4_math) curated by the OpenThoughts4 team. The dataset provides responses from [Qwen/Qwen3-235B-A22B-FP8](https://huggingface.co/Qwen/Qwen3-235B-A22B-FP8) with generation parameters of temperature = 0.8 and max output tokens = 16000. The dataset structure includes columns such as instruction_seed (original math problem/question text without chat formatting), _source (the origin dataset), gpt41_mini_response (a reference solution generated by GPT-4.1 Mini), __original_row_idx (the row index from the original source dataset), length (the token count of the response), ms_id (a unique sample identifier), generated_text (a response including chain-of-thought with <think> tags, generated by Qwen3-32B from the parent dataset), qwen235b_generated_text (a response including chain-of-thought with <think> tags, generated by Qwen3-235B-A22B-FP8), and conversations (the prompt and response in chat format). The main difference from the parent dataset is the addition of the qwen235b_generated_text column, which contains responses generated by Qwen3-235B-A22B-FP8.