marin-community/open-thoughts-4-128-math-qwen3-32b-annotated-32768-tokens

--- dataset_info: features: - name: row_id dtype: int64 - name: instruction_seed dtype: string - name: _source dtype: string - name: gpt41_mini_response dtype: string - name: __original_row_idx dtype: int64 - name: length dtype: int64 - name: ms_id dtype: int64 - name: generated_text dtype: string - name: final_answer dtype: string - name: complete_responses_count dtype: int64 splits: - name: train num_bytes: 60577594 num_examples: 1024 download_size: 18689160 dataset_size: 60577594 configs: - config_name: default data_files: - split: train path: data/train-* --- # open-thoughts-4-128-math-qwen3-32b-annotated-32768-tokens Math reasoning responses generated by **Qwen3-32B** ([Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B)). ## Overview - **Total rows:** 1,024 - **Unique prompts:** 128 (each with 8 response annotations) - **Source prompts:** [marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted](https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted) - **Prompt alignment:** Exact `instruction_seed` match to [marin-community/open-thoughts-4-128-math-kimi-k2pt5-annotated-32768-tokens](https://huggingface.co/datasets/marin-community/open-thoughts-4-128-math-kimi-k2pt5-annotated-32768-tokens) - **Generation model:** [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) - **Max tokens:** 32,768 - **Temperature:** 0.8 - **Tokenizer used for stats:** Qwen/Qwen2.5-3B ## Statistics | Metric | Value | |--------|-------| | Avg tokens per response | 17,088 | | Median tokens per response | 14,557 | | Responses with `<think>` tag | 1020/1024 (99.6%) | | Complete responses (has `</think>` + `\boxed{...}`) | 910/1024 (88.9%) | | Truncated responses | 110/1024 (10.7%) | | Empty responses | 4/1024 (0.4%) | ## Columns | Column | Description | |--------|-------------| | `row_id` | Row identifier preserved from the source dataset | | `instruction_seed` | The math problem prompt | | `generated_text` | Qwen3-32B generated response with a `<think>...</think>` reasoning trace when present | | `ms_id` | Math seed ID, groups all 8 responses for the same prompt | | `_source` | Source dataset identifier | | `gpt41_mini_response` | GPT-4.1 mini reference response | | `__original_row_idx` | Row index from the pre-reformatted source pipeline | | `length` | Length metadata carried over from the source dataset | | `final_answer` | Extracted final answer when present | | `complete_responses_count` | Number of complete responses in the source n=8 group for the prompt | ## Response Format Each response in the `generated_text` column generally follows this format: ```text <think> [model reasoning trace] </think> [final answer, typically containing \boxed{...}] ``` Most responses include an opening `<think>` tag; empty responses account for the remainder. Responses that are truncated may be missing the closing `</think>` tag and or the `\boxed{...}` answer. ## Construction Created by taking the first 1,024 rows of [marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted](https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted) without shuffling. The `instruction_seed` sequence was checked against the Kimi K2.5 128-prompt reference dataset and matched exactly across all 1,024 rows.数据集信息： 特征： - 字段名：row_id，数据类型：int64 - 字段名：instruction_seed，数据类型：string - 字段名：_source，数据类型：string - 字段名：gpt41_mini_response，数据类型：string - 字段名：__original_row_idx，数据类型：int64 - 字段名：length，数据类型：int64 - 字段名：ms_id，数据类型：int64 - 字段名：generated_text，数据类型：string - 字段名：final_answer，数据类型：string - 字段名：complete_responses_count，数据类型：int64 划分集： - 划分名：train，字节数：60577594，样本数：1024 下载大小：18689160，数据集总大小：60577594 配置： - 配置名：default，数据文件： - 划分：train，路径：data/train-* --- # open-thoughts-4-128-math-qwen3-32b-annotated-32768-tokens 由**Qwen3-32B** ([Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B)) 生成的数学推理响应。 ## 概览 - **总样本数**：1024 - **唯一提示词**：128个（每个对应8个响应标注） - **源提示词数据集**：[marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted](https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted) - **提示词对齐**：与[marin-community/open-thoughts-4-128-math-kimi-k2pt5-annotated-32768-tokens](https://huggingface.co/datasets/marin-community/open-thoughts-4-128-math-kimi-k2pt5-annotated-32768-tokens)的`instruction_seed`完全匹配 - **生成模型**：[Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) - **最大Token数**：32768 - **温度系数（Temperature）**：0.8 - **用于统计的分词器**：Qwen/Qwen2.5-3B ## 统计指标 | 指标 | 数值 | |--------|-------| | 单响应平均Token数 | 17088 | | 单响应Token数中位数 | 14557 | | 包含`<think>`标签的响应 | 1020/1024（99.6%） | | 完整响应（同时包含`</think>`与`oxed{...}`） | 910/1024（88.9%） | | 截断响应 | 110/1024（10.7%） | | 空响应 | 4/1024（0.4%） | ## 字段说明 | 字段 | 描述 | |--------|-------------| | `row_id` | 源自源数据集的行标识符 | | `instruction_seed` | 数学问题提示词 | | `generated_text` | Qwen3-32B生成的响应，若存在则包含`<think>...</think>`形式的推理轨迹 | | `ms_id` | 数学种子ID，用于将同一提示词对应的8个响应归为一组 | | `_source` | 源数据集标识符 | | `gpt41_mini_response` | GPT-4.1 mini 参考响应 | | `__original_row_idx` | 预格式化前的源数据流水线中的行索引 | | `length` | 源自源数据集的长度元数据 | | `final_answer` | 提取得到的最终答案（若存在） | | `complete_responses_count` | 该提示词对应的n=8组源数据中完整响应的数量 | ## 响应格式 `generated_text`列中的每个响应通常遵循以下格式： text <think> [模型推理轨迹] </think> [最终答案，通常包含oxed{...}] 大多数响应包含开头的`<think>`标签，剩余为少量空响应。截断的响应可能缺失闭合的`</think>`标签或`oxed{...}`形式的答案。 ## 数据集构建 本数据集通过未进行洗牌操作，直接截取[marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted](https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted)的前1024行构建而成。 我们将`instruction_seed`序列与Kimi K2.5的128提示词参考数据集进行比对，确认1024条样本的`instruction_seed`完全匹配。