IntelligenceLab/Cos-Play-Cold-Start

---
license: mit
task_categories:
- reinforcement-learning
language:
- en
tags:
- game-playing
- llm-agent
- cold-start
- skill-labeling
- grpo
pretty_name: "COS-PLAY Cold-Start Data"
size_categories:
- 10K<n<100K
configs:
- config_name: episodes_twenty_forty_eight
data_files: "data/episodes/twenty_forty_eight.jsonl"
- config_name: episodes_tetris
data_files: "data/episodes/tetris.jsonl"
- config_name: episodes_candy_crush
data_files: "data/episodes/candy_crush.jsonl"
- config_name: episodes_super_mario
data_files: "data/episodes/super_mario.jsonl"
- config_name: episodes_sokoban
data_files: "data/episodes/sokoban.jsonl"
- config_name: episodes_pokemon_red
data_files: "data/episodes/pokemon_red.jsonl"
- config_name: episodes_avalon
data_files: "data/episodes/avalon.jsonl"
- config_name: episodes_diplomacy
data_files: "data/episodes/diplomacy.jsonl"
- config_name: grpo_action_taking_twenty_forty_eight
data_files: "data/grpo_coldstart/twenty_forty_eight/action_taking.jsonl"
- config_name: grpo_action_taking_tetris
data_files: "data/grpo_coldstart/tetris/action_taking.jsonl"
- config_name: grpo_action_taking_candy_crush
data_files: "data/grpo_coldstart/candy_crush/action_taking.jsonl"
- config_name: grpo_action_taking_super_mario
data_files: "data/grpo_coldstart/super_mario/action_taking.jsonl"
- config_name: grpo_action_taking_sokoban
data_files: "data/grpo_coldstart/sokoban/action_taking.jsonl"
- config_name: grpo_action_taking_pokemon_red
data_files: "data/grpo_coldstart/pokemon_red/action_taking.jsonl"
- config_name: grpo_skill_selection_twenty_forty_eight
data_files: "data/grpo_coldstart/twenty_forty_eight/skill_selection.jsonl"
- config_name: grpo_skill_selection_tetris
data_files: "data/grpo_coldstart/tetris/skill_selection.jsonl"
- config_name: grpo_skill_selection_candy_crush
data_files: "data/grpo_coldstart/candy_crush/skill_selection.jsonl"
- config_name: grpo_skill_selection_super_mario
data_files: "data/grpo_coldstart/super_mario/skill_selection.jsonl"
- config_name: grpo_skill_selection_sokoban
data_files: "data/grpo_coldstart/sokoban/skill_selection.jsonl"
- config_name: grpo_skill_selection_pokemon_red
data_files: "data/grpo_coldstart/pokemon_red/skill_selection.jsonl"
---

# COS-PLAY Cold-Start Data

Pre-generated cold-start data for [COS-PLAY](https://github.com/wuxiyang1996/cos-play) (COLM 2026): **Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Game Play**.

## Dataset Summary

This dataset contains GPT-5.4-generated seed trajectories and skill-labeled episodes for 8 games, used to bootstrap the COS-PLAY co-evolution training loop.

| Game | Episodes | Steps (action) | Steps (skill) |
|------|----------|-----------------|----------------|
| 2048 | 60 | 8,125 | varies |
| Tetris | 60 | 3,700 | varies |
| Candy Crush | 60 | 3,000 | varies |
| Super Mario | 60 | 3,043 | varies |
| Sokoban | 59 | 5,204 | varies |
| Pokemon Red | 60 | 11,552 | varies |
| Avalon | 60 | — | — |
| Diplomacy | 60 | — | — |

## Dataset Structure

### Episodes (`data/episodes/<game>.jsonl`)

Each line is a full episode with fields:

- `episode_id` — unique episode identifier
- `game_name` — game name
- `experiences` — list of step-level data, each containing:
- `state`, `action`, `reward`, `next_state`, `done`
- `summary_state` — structured state summary
- `intentions` — agent's declared intention at the step
- `available_actions` — list of legal actions

### GRPO Cold-Start (`data/grpo_coldstart/<game>/`)

Training data for GRPO LoRA fine-tuning of the decision agent:

- **`action_taking.jsonl`** — one row per step: state + actions → chosen action
- **`skill_selection.jsonl`** — one row per step with ≥2 skill candidates: state + candidates → chosen skill

Fields: `type`, `game`, `episode`, `step`, `prompt`, `chosen`, `rejected`

## Usage

### Download with Python

```python
from huggingface_hub import snapshot_download

snapshot_download(
repo_id="IntelligenceLab/Cos-Play-Cold-Start",
repo_type="dataset",
local_dir="labeling/output/gpt54_skill_labeled",
)
```

### Download with CLI

```bash
pip install huggingface_hub
huggingface-cli download IntelligenceLab/Cos-Play-Cold-Start \
--repo-type dataset \
--local-dir labeling/output/gpt54_skill_labeled
```

### Load with `datasets`

```python
from datasets import load_dataset

# Load episodes for a specific game
ds = load_dataset("IntelligenceLab/Cos-Play-Cold-Start", "episodes_tetris")

# Load GRPO action-taking data
ds = load_dataset("IntelligenceLab/Cos-Play-Cold-Start", "grpo_action_taking_tetris")
```

## Citation

```bibtex
@inproceedings{cosplay2026,
title={COS-PLAY: Co-Evolving LLM Decision and Skill Bank Agents for Long-Horizon Game Play},
author={...},
booktitle={COLM},
year={2026}
}
```