marin-community/open-thoughts-4-8530-math-kimi-k2pt5-annotated-32768-tokens

---
dataset_info:
split: train
num_examples: 68240
license: apache-2.0
---

# open-thoughts-4-8530-math-kimi-k2pt5-annotated-32768-tokens

This dataset contains 68240 rows (8530 math prompts x 8 responses each) generated by
**Kimi K2.5** with a max token length of 32768.

## Columns

| Column | Description |
|--------|-------------|
| `row_id` | Original row identifier |
| `instruction_seed` | The math prompt |
| `_source` | Source dataset for the prompt |
| `gpt41_mini_response` | GPT-4.1-mini reference response |
| `__original_row_idx` | Original row index in source |
| `length` | Prompt length |
| `ms_id` | Source ID |
| `kimi_k2pt5_generated_text` | Response generated by Kimi K2.5 (the primary response column) |
| `final_answer` | Last `\boxed{...}` answer extracted from `kimi_k2pt5_generated_text` after `</think>` |
| `complete_responses_count` | Number of the 8 responses per prompt that are "complete" (have `</think>` followed by `\boxed{...}`) |

## Stats

- **Complete responses**: 66368 rows have at least one complete sibling
- **Prompts with all 8 complete**: 4810 / 8530
- **Prompts with 0 complete**: 234 / 8530

## Purpose

This dataset shares the same 8530 prompts as
[marin-community/open-thoughts-4-8530-math-qwen3-30b-a3B-thinking-2507-annotated-32768-tokens](https://huggingface.co/datasets/marin-community/open-thoughts-4-8530-math-qwen3-30b-a3B-thinking-2507-annotated-32768-tokens),
enabling direct comparison of Kimi K2.5 vs Qwen3-30B-A3B responses on the same math problems.

---
数据集信息：
拆分集：训练集（train）
样本总数：68240
许可证：Apache-2.0
---

# open-thoughts-4-8530-math-kimi-k2pt5-annotated-32768-tokens

本数据集共包含68240条数据（即8530道数学提示词，每道提示词对应8条生成回复），由**Kimi K2.5**生成，单条文本最大Token长度为32768。

## 字段说明

| 字段名 | 描述 |
|--------|-------------|
| `row_id` | 原始行标识符 |
| `instruction_seed` | 数学提示词（prompt） |
| `_source` | 该提示词的源数据集 |
| `gpt41_mini_response` | GPT-4.1-mini 参考回复 |
| `__original_row_idx` | 源数据集中的原始行索引 |
| `length` | 提示词长度 |
| `ms_id` | 源数据标识 |
| `kimi_k2pt5_generated_text` | Kimi K2.5 生成的回复（核心回复字段） |
| `final_answer` | 从`kimi_k2pt5_generated_text`的`</think>`标记后提取的最后一个`oxed{...}`格式答案 |
| `complete_responses_count` | 单提示词的8条回复中「完整回复」的数量，完整回复指包含`</think>`后接`oxed{...}`格式内容的回复 |

## 统计信息

- **完整回复覆盖**：66368条数据至少拥有1条完整的同提示词回复
- **8条回复均完整的提示词**：4810 / 8530
- **无完整回复的提示词**：234 / 8530

## 数据集用途

本数据集与[marin-community/open-thoughts-4-8530-math-qwen3-30b-a3B-thinking-2507-annotated-32768-tokens](https://huggingface.co/datasets/marin-community/open-thoughts-4-8530-math-qwen3-30b-a3B-thinking-2507-annotated-32768-tokens)使用完全相同的8530道数学提示词，可直接对比Kimi K2.5与Qwen3-30B-A3B在同一数学题目上的生成回复效果。