Hugging Face's logo
Activity Feed

AI & ML interests

Autoregressive Models, Transformers, Reinforcement Learning, Offline RL, Continuous Control, MuJoCo, Unity ML-Agents, Discrete Control, Multi-Agent RL

Recent Activity
kissin42  updated a model 3 days ago
ccnets/causal-gpt-rl-unity
kissin42  updated a model 6 days ago
ccnets/causal-gpt-rl
kissin42  updated a dataset 6 days ago
ccnets/causal-gpt-rl-datasets
View all activity

Organization Card
Community About org cards

CCNets, Inc.

Causal GPT-RL — GPT-style transformers (Llama) running as RL policies. The same architecture covers continuous motor control, discrete goal games, hybrid action spaces, and cooperative or competitive multi-agent scenes — one model design for every action-space type.

Both LLM generation and RL interaction are autoregressive:

token           → next token                           (LLM generation)
(state, action) → (next state from env, next action)   (RL rollout)

Running that loop on its own outputs is where transformers have historically drifted. These policies remain stable over full episodes and well beyond their trained context window, with rollout history set at load time rather than through retraining.

From environment to dataset to policy

Four repositories, one chain:
environment → recorded trajectories → offline RL training → policy bundle

Repo Link in the chain
causal-gpt-rl-unity-envs The release_23 builds, model-removed — where the trajectories come from.
causal-gpt-rl-unity-datasets The trajectories recorded from them — every environment in three calibrated tiers.
causal-gpt-rl-unity The policies trained on those — safetensors plus per-context ONNX for Sentis.
causal-gpt-rl The MuJoCo policies — the same chain on Gymnasium and Minari simple/medium. No expert data.

The policies span several action spaces and head types — continuous Box, Discrete and MultiDiscrete — across single-agent, cooperative and competitive scenes, from Unity ML-Agents to the MuJoCo continuous-control tasks. Each is trained on the trajectories above: the Unity tiers from the dataset repo, the MuJoCo ones from Minari. All of them are the same GPT-style backbone, the same configuration, the same bundle format. MuJoCo training runs: wandb.ai/causal-gpt-rl/mujoco.

Licensing

Three different terms apply, by artifact type:

  • Models and code — PolyForm Noncommercial 1.0.0
  • Trajectory data (causal-gpt-rl-unity-datasets) — CC-BY-4.0. Please credit ccnets — Causal GPT-RL and note the Unity ML-Agents source environment.
  • Unity builds and stock policies (causal-gpt-rl-unity-envs) — Apache-2.0, as derivatives of Unity ML-Agents material.

Links