---
license: other
license_name: polyform-noncommercial-1.0.0
license_link: https://polyformproject.org/licenses/noncommercial/1.0.0/
library_name: safetensors
tags:
- reinforcement-learning
- offline-rl
- mujoco
- gpt
- llama
- autoregressive
- causal-gpt-rl
---

# Causal GPT-RL

**First decoder-only transformer (GPT) to reach expert level on Humanoid offline RL from medium data — beyond what Behavior Cloning and Decision Transformer can achieve.**

GPT-style transformers (GPT-2, Llama) running as RL policies in continuous-control environments.

```text
action → next state → next action      (RL rollouts)
token  → next token  → next token      (LLM generation)
```

Stable under self-generated rollouts — long-horizon control without the drift that has historically kept transformers from being usable as RL agents.

## Bundles in this repository

| Environment | Subfolder | Context length | Return (mean ± std) |
|---|---|---|---|
| Ant-v5 | `ant-v5` | 16 | 2614 ± 1515 |
| HalfCheetah-v5 | `halfcheetah-v5` | 32 | 3251 ± 1916 |
| Walker2d-v5 | `walker2d-v5` | 24 | 2345 ± 879 |
| Humanoid-v5 | `humanoid-v5` | 32 | 2371 ± 2850 |

Returns are over 5 episodes with `seed=0` (HalfCheetah-v5: 50 episodes), run on CPU via `run_episodes`.

## Quick Start

```bash
pip install "causal-gpt-rl[hub,mujoco]"
```

```python
import gymnasium as gym
from causal_gpt_rl.inference import load_runner_from_hub, run_episodes

env = gym.make("Ant-v5")
runner = load_runner_from_hub(
    repo_id="ccnets/causal-gpt-rl",
    subfolder="ant-v5",
)
stats = run_episodes(env, runner, num_episodes=5, seed=0)
print(stats["return_mean"], stats["return_std"])
```

## Bundle contents

Each subfolder contains:

- `model.safetensors` — model state dict for inference
- `config.json` — model config, observation specs, action specs, context length
- `state_normalizer.safetensors` — state normalization statistics

## Model details

Llama-style transformer decoder, 4 layers, 8 heads. Hidden size 192 for Ant/HalfCheetah/Walker2d, 256 for Humanoid.

## Training data

[Minari](https://minari.farama.org/) `mujoco/{env}/simple-v0` + `mujoco/{env}/medium-v0` per environment (expert split not used).

## Links

- **Code:** [github.com/ccnets-team/causal-gpt-rl](https://github.com/ccnets-team/causal-gpt-rl)
- **Training logs (W&B):** [wandb.ai/junhopark/Causal GPT-RL](https://wandb.ai/junhopark/Causal%20GPT-RL)
- **Website:** [ccnets.org](https://ccnets.org)

## License

PolyForm Noncommercial License 1.0.0. For commercial use, contact via [ccnets.org](https://ccnets.org).