AntimLabs/FlappyBird-SFT

---
license: mit
task_categories:
- reinforcement-learning
- text-generation
tags:
- flappy-bird
- reinforcement-learning
- game-ai
- sft
---

# FlappyBird SFT Dataset

This dataset contains supervised fine-tuning (SFT) data for Flappy Bird game control policy.

## Dataset Details

- **Source**: Evaluation results from `AntimLabs/Qwen2.5-1.5B-Instruct-FlappyBird-RL-71`
- **Format**: Standard SFT format with `prompt` and `completion` columns
- **Examples**: 710 game episodes
- **Total conversation turns**: ~147k

## Data Format

Each example has two columns:

- **prompt**: List of messages containing system prompt and initial user observation
- **completion**: List of messages containing the full game conversation (user observations and assistant actions)

```python
{
"prompt": [
{"role": "system", "content": "You are the Flappy Bird control policy..."},
{"role": "user", "content": "<FLAPPY id=1>..."}
],
"completion": [
{"role": "assistant", "content": "<ACTIONS>[]</ACTIONS>"},
{"role": "user", "content": "<FLAPPY id=2>..."},
{"role": "assistant", "content": "<ACTIONS>[TAP]</ACTIONS>"},
...
]
}
```

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("AntimLabs/FlappyBird-SFT", split="train")
```

license: MIT协议
task_categories:
- 强化学习（Reinforcement Learning）
- 文本生成
tags:
- Flappy Bird
- 强化学习（Reinforcement Learning）
- 游戏AI（Game AI）
- 监督微调（Supervised Fine-Tuning，SFT）

# Flappy Bird 监督微调数据集

本数据集包含用于Flappy Bird游戏控制策略的监督微调（Supervised Fine-Tuning，SFT）数据。

## 数据集详情

- **数据来源**：来自`AntimLabs/Qwen2.5-1.5B-Instruct-FlappyBird-RL-71`的评估结果
- **数据格式**：采用标准监督微调格式，包含`prompt`与`completion`两列
- **样本数量**：710局游戏对局
- **总对话轮次**：约14.7万

## 数据格式

每个样本包含两列：

- **prompt**：包含系统提示与初始用户观测信息的消息列表
- **completion**：包含完整游戏对话（用户观测与智能体动作）的消息列表

python
{
"prompt": [
{"role": "system", "content": "You are the Flappy Bird control policy..."},
{"role": "user", "content": "<FLAPPY id=1>..."}
],
"completion": [
{"role": "assistant", "content": "<ACTIONS>[]</ACTIONS>"},
{"role": "user", "content": "<FLAPPY id=2>..."},
{"role": "assistant", "content": "<ACTIONS>[TAP]</ACTIONS>"},
...
]
}

## 使用方法

python
from datasets import load_dataset

dataset = load_dataset("AntimLabs/FlappyBird-SFT", split="train")