tsilva/gymrec__BreakoutNoFrameskip_dash_v4
收藏Hugging Face2026-03-10 更新2026-03-29 收录
下载链接:
https://hf-mirror.com/datasets/tsilva/gymrec__BreakoutNoFrameskip_dash_v4
下载链接
链接失效反馈官方服务:
资源简介:
---
language: en
license: mit
size_categories:
- 100K<n<1M
task_categories:
- reinforcement-learning
pretty_name: BreakoutNoFrameskip-v4 Gameplay Dataset
tags:
- gymnasium
- atari
- BreakoutNoFrameskip-v4
---
# BreakoutNoFrameskip-v4 Gameplay Dataset
Gameplay recordings (collected by: `breakout`) from the Gymnasium environment `BreakoutNoFrameskip-v4`,
captured using [gymrec](https://github.com/tsilva/gymrec).
## Dataset Summary
| Stat | Value |
|------|-------|
| Total frames | 494,029 |
| Episodes | 50 |
| Environment | `BreakoutNoFrameskip-v4` |
| Backend | Atari (ALE-py) |
| Collector(s) | breakout |
| gymrec version(s) | 0.1.0+23e91c8, 0.1.0+72aad18 |
## Environment Configuration
| Setting | Value |
|---------|-------|
| Frameskip | 1 |
| Target FPS | 30 |
| Sticky Actions | 0.0 |
| Observation Shape | 210 × 160 × 3 |
| Observation Dtype | uint8 |
| Action Space | Discrete |
| Number of Actions | 4 |
## Dataset Structure
Minari-compatible flat table format. Use `minari-export` for native [Minari](https://minari.farama.org/) HDF5 format.
Each episode has N step rows plus one terminal observation row (N+1 pattern).
The terminal observation is the final state after the last step — it has an empty action
and null values for rewards/terminations/truncations/infos.
- **episode_id** (`binary(16)`): Unique UUID identifier for each episode (16 bytes, universally unique across all recordings)
- **seed** (`int` or `null`): RNG seed used for `env.reset()` (set on first row of each episode, `null` on other rows)
- **observations** (`Image`): RGB frame from the environment
- **actions** (`list`): Action taken at this step (`[]` for terminal observations)
- **rewards** (`float` or `null`): Reward received (`null` on terminal observation rows)
- **terminations** (`bool` or `null`): Whether the episode terminated naturally (`null` on terminal observation rows)
- **truncations** (`bool` or `null`): Whether the episode was truncated (`null` on terminal observation rows)
- **infos** (`str` or `null`): Additional environment info as JSON (`null` on terminal observation rows)
- **session_id** (`binary(16)`): UUID grouping all episodes from one `gymrec record` run
- **collector** (`string`): Who collected the data (`"human"`, `"random"`, or future agent names)
- **gymrec_version** (`string`): Version of gymrec used to record (e.g. `"0.1.0+abc1234"`)
## Usage
```python
from datasets import load_dataset
ds = load_dataset("tsilva/gymrec__BreakoutNoFrameskip_dash_v4")
```
## About
Recorded with [gymrec](https://github.com/tsilva/gymrec).
Curated by: tsilva
提供机构:
tsilva



