reasoning-degeneration-dev/wmc-sft-baseline-v2

---
license: mit
tags:
- world-model-curiosity
- sft-warmup
- baseline
- countdown
---

# wmc-sft-baseline-v2

SFT warmup dataset for baseline GRPO condition. 1200 Countdown problems (5 numbers, +/-/*) solved by Qwen3-1.7B with 32k token generation. Contains full untruncated reasoning traces. No confidence tags.

## Dataset Info

- **Rows**: 1200
- **Columns**: 3

## Columns

| Column | Type | Description |
|--------|------|-------------|
| messages | List({'content': Value('string'), 'role': Value('string')}) | Chat-format conversation (system, user, assistant). System sets task format, user provides numbers/target, assistant contains full reasoning trace in <think> tags followed by answer. |
| correct | Value('bool') | Boolean indicating whether the model's final expression evaluates to the target number. |
| difficulty | Value('string') | Problem difficulty tier: easy, medium, or hard (based on number count and operator complexity). |

## Generation Parameters

```json
{
"script_name": "generate_sft_data.py",
"model": "Qwen/Qwen3-1.7B",
"description": "SFT warmup dataset for baseline GRPO condition. 1200 Countdown problems (5 numbers, +/-/*) solved by Qwen3-1.7B with 32k token generation. Contains full untruncated reasoning traces. No confidence tags.",
"hyperparameters": {
"temperature": 0.7,
"max_tokens": 32768,
"top_p": 0.9
},
"input_datasets": []
}
```

## Experiment Documentation

For complete experiment details, see [https://github.com/Zayne-sprague/SC-Research-Notes/tree/main/experiments/world-model-curiosity](https://github.com/Zayne-sprague/SC-Research-Notes/tree/main/experiments/world-model-curiosity)

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("reasoning-degeneration-dev/wmc-sft-baseline-v2", split="train")
print(f"Loaded {len(dataset)} rows")
```

---

*This dataset is tracked in [reasoning-degeneration-dev/PROJECT-MANIFEST](https://huggingface.co/datasets/reasoning-degeneration-dev/PROJECT-MANIFEST)*

许可证：MIT许可证
标签：
- 世界模型好奇心（world-model-curiosity）
- 监督微调预热（Supervised Fine-Tuning, SFT）
- 基线
- 倒计时

# wmc-sft-baseline-v2

本数据集为基线GRPO条件下的监督微调预热数据集，包含1200个倒计时问题（使用5个数字与加减乘运算符），由通义千问3-1.7B（Qwen3-1.7B）模型生成，生成长度上限为32768词元（Token），包含完整未截断的推理轨迹，无置信度标签。

## 数据集信息
- **行数**: 1200
- **列数**: 3

## 列信息

| 列名 | 数据类型 | 描述 |
|------|----------|------|
| messages | List({'content': Value('string'), 'role': Value('string')}) | 聊天格式对话（包含系统、用户、助手三类角色）：系统提示用于设定任务格式，用户输入数字与目标数值，助手回复则包含包裹在<think>标签内的完整推理轨迹，其后跟随最终答案。 |
| correct | Value('bool') | 布尔值，用于指示模型生成的最终表达式计算结果是否与目标数值一致。 |
| difficulty | Value('string') | 字符串类型的问题难度层级，分为简单（easy）、中等（medium）、困难（hard）三类，划分依据为数字数量与运算符复杂度。 |

## 生成参数
json
{
"script_name": "generate_sft_data.py",
"model": "Qwen/Qwen3-1.7B",
"description": "本数据集为基线GRPO条件下的监督微调预热数据集，包含1200个倒计时问题（使用5个数字与加减乘运算符），由通义千问3-1.7B（Qwen3-1.7B）模型生成，生成长度上限为32768词元（Token），包含完整未截断的推理轨迹，无置信度标签。",
"hyperparameters": {
"temperature": 0.7,
"max_tokens": 32768,
"top_p": 0.9
},
"input_datasets": []
}

## 实验文档
如需查看完整实验细节，请访问：[https://github.com/Zayne-sprague/SC-Research-Notes/tree/main/experiments/world-model-curiosity](https://github.com/Zayne-sprague/SC-Research-Notes/tree/main/experiments/world-model-curiosity)

## 使用方法
python
from datasets import load_dataset

dataset = load_dataset("reasoning-degeneration-dev/wmc-sft-baseline-v2", split="train")
print(f"已加载 {len(dataset)} 条数据")

*本数据集已在[reasoning-degeneration-dev/PROJECT-MANIFEST](https://huggingface.co/datasets/reasoning-degeneration-dev/PROJECT-MANIFEST)中进行追踪*