---
license: apache-2.0
dataset_info:
  features:
  - name: observation
    dtype:
      array3_d:
        shape:
        - 64
        - 64
        - 3
        dtype: uint8
  - name: action
    dtype: uint8
  - name: reward
    dtype: float32
  - name: done
    dtype: bool
  - name: truncated
    dtype: bool
  splits:
  - name: train
    num_bytes: 26043525000
    num_examples: 900000
  - name: test
    num_bytes: 2893725000
    num_examples: 100000
  download_size: 3128341675
  dataset_size: 28937250000
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
task_categories:
- reinforcement-learning
language:
- en
tags:
- procgen
- bigfish
- benchmark
- openai
pretty_name: Procgen Benchmark - Bigfish
size_categories:
- 100K<n<1M
---
# Procgen Benchmark - Bigfish

<video controls autoplay src="https://cdn-uploads.huggingface.co/production/uploads/633c1daf31c06121a58f2df9/brMaX1xgew7ulqkMU0Ahi.mp4"></video>

This dataset contains trajectories generated by a [PPO](https://arxiv.org/abs/1707.06347) reinforcement learning agent trained on the Bigfish environment from the [Procgen Benchmark](https://openai.com/index/procgen-benchmark/). The agent has been trained for 50M steps and the final evaluation performance is `32.33`.

## Dataset Usage

Regular usage:
```python
from datasets import load_dataset
train_dataset = load_dataset("EpicPinkPenguin/procgen_bigfish", split="train")
test_dataset = load_dataset("EpicPinkPenguin/procgen_bigfish", split="test")
```

Usage with PyTorch:
```python
from datasets import load_dataset
train_dataset = load_dataset("EpicPinkPenguin/procgen_bigfish", split="train").with_format("torch")
test_dataset = load_dataset("EpicPinkPenguin/procgen_bigfish", split="test").with_format("torch")
```

## Dataset Structure
### Data Instances
Each data instance represents a single step consisting of tuples of the form (observation, action, reward, done, truncated) = (o_t, a_t, r_{t+1}, done_{t+1}, trunc_{t+1}).

```json
{'action': 1,
 'done': False,
 'observation': [[[0, 166, 253],
                  [0, 174, 255],
                  [0, 170, 251],
                  [0, 191, 255],
                  [0, 191, 255],
                  [0, 221, 255],
                  [0, 243, 255],
                  [0, 248, 255],
                  [0, 243, 255],
                  [10, 239, 255],
                  [25, 255, 255],
                  [0, 241, 255],
                  [0, 235, 255],
                  [17, 240, 255],
                  [10, 243, 255],
                  [27, 253, 255],
                  [39, 255, 255],
                  [58, 255, 255],
                  [85, 255, 255],
                  [111, 255, 255],
                  [135, 255, 255],
                  [151, 255, 255],
                  [173, 255, 255],
...
                  [0, 0, 37],
                  [0, 0, 39]]],
 'reward': 0.0,
 'truncated': False}
```

### Data Fields
- `observation`: The current RGB observation from the environment.
- `action`: The action predicted by the agent for the current observation.
- `reward`: The received reward from stepping the environment with the current action.
- `done`: If the new observation is the start of a new episode. Obtained after stepping the environment with the current action.
- `truncated`: If the new observation is the start of a new episode due to truncation. Obtained after stepping the environment with the current action.

### Data Splits
The dataset is divided into a `train` (90%) and `test` (10%) split

## Dataset Creation
The dataset was created by training an RL agent with [PPO](https://arxiv.org/abs/1707.06347) for 50M steps on the Procgen Bigfish environment. The agent obtained a final performance of `32.33`. The trajectories where generated by taking the argmax action at each step, corresponding to taking the mode of the action distribution.

## Procgen Benchmark
The [Procgen Benchmark](https://openai.com/index/procgen-benchmark/), released by OpenAI, consists of 16 procedurally-generated environments designed to measure how quickly reinforcement learning (RL) agents learn generalizable skills. It emphasizes experimental convenience, high diversity within and across environments, and is ideal for evaluating both sample efficiency and generalization. The benchmark allows for distinct training and test sets in each environment, making it a standard research platform for the OpenAI RL team. It aims to address the need for more diverse RL benchmarks compared to complex environments like Dota and StarCraft.