ccnets/causal-gpt-rl-datasets

Causal GPT-RL — MuJoCo trajectories

MuJoCo trajectories packaged as Minari datasets, in offline-RL form. What makes this repository different from a collection of rollouts is where the trajectories come from: they were recorded by a policy that had never seen expert data.

The first dataset is 1000 episodes of Humanoid-v5, recorded by a policy trained on simple and medium trajectories alone — one episode at every retention length from 1 to 1000, each value used exactly once.

Why this repository exists

The scarcity of good trajectories is the binding constraint in offline RL. There is no internet-scale corpus of behaviour, and the usual way out is circular — expert datasets need an expert policy, and an expert policy is most of the problem already solved.

So the question is not only can a model learn from the current dataset, but can it produce a better next one:

simple                     ->  model  ->  medium
simple + medium            ->  model  ->  better than medium
simple + medium + that     ->  model  ->  expert

Each dataset here is one turn of that loop. No expert trajectories are used at any point.

Where the loop stands

Normalized skill on the model card's scale — a random policy is 0, expert is 100:

	dataset	mean return	normalized
input	Minari mujoco/humanoid/simple-v0	5484.6	63.3
input	Minari mujoco/humanoid/medium-v0	7014.5	81.3
this repo	ccnets/humanoid/medium-v0	7494.3	86.9
target	expert	8605.1	100.0

"Minari" marks the two Farama-published datasets this project trained on. Everything else in this repository is recorded, not downloaded.

The recorded trajectories score +479.8 return above the medium dataset the policy was trained on, and 85.7 % of the 1000 episodes beat that dataset's mean. The gap left to expert is 13.1 normalized points.

humanoid/medium-v0

Recorded by the ccnets/causal-gpt-rl/humanoid-v5 bundle, trained on the Minari simple-v0 and medium-v0 trajectories of that environment with no expert data.

Episodes	1000
Transitions	924,747
Episode length	924.7 mean (87 min, 1000 max)
Return	7494.3 ± 1667.6 (515.4 min, 8340.7 max)
Ended by termination	156
Ended by the 1000-step limit	844
Observation / action	Box(348,) / Box(17,), float32

One retention per episode

Every episode carries a distinct seed and a different retention — kv_cache_max_len from 1 to 1000, each value used exactly once. Retention is how many steps of its own past the policy keeps while acting. It is a load-time argument, so all 1000 episodes come from the same weights and the same trained context length of 32; only how much history each rollout kept differs.

episode_metadata.jsonl maps every episode file to the retention that recorded it — episodes are numbered in the order they finished, not by retention:

{"episode": "ep_000000", "row": 4, "transitions": 113, "terminated": true, "truncated": false, "kv_cache_max_len": 16}

That makes the dataset usable two ways: as offline-RL trajectories, and as a measurement of what retention does to a policy's behaviour. Within it, episodes recorded between 16 and 128 steps of retention average 91.1 normalized against 85.5 for those above 500 — but with one episode per retention, seed variance is larger than the effect, so treat the per-retention axis as evidence to build on rather than a settled result.

Files

humanoid/medium-v0/
  data/main_data.hdf5        the Minari dataset
  data/metadata.json
  episode_metadata.jsonl     one line per episode: file, row, length, retention
  collection_spec.json       what recorded it: bundle, context length, retentions
  validation_summary.json    every number quoted above

Load it

import minari

dataset = minari.load_dataset("ccnets/humanoid/medium-v0")
episode = next(dataset.iterate_episodes())
print(episode.observations.shape, episode.rewards.sum())

Place the humanoid/medium-v0 directory under ccnets/ in your Minari datasets path — ~/.minari/datasets/ccnets/humanoid/medium-v0 — or point MINARI_DATASETS_PATH at wherever you keep it.

The files were recorded with torch 2.12.0, gymnasium 1.2.2 and mujoco 3.8.1; a different simulator release is a different measurement even with identical weights and seeds. The recorder and packager are collection/, and the cycle this repository is built on is described in Improving the next dataset.

A note on the name

ccnets/humanoid/medium-v0 keeps the rung — medium-v0 — and changes the namespace. The rung says where in the ladder the data was produced: after the medium trajectories, before expert. The namespace says who produced it, and keeps it clear of Farama's mujoco/humanoid/medium-v0, which is the dataset the recording policy was trained on. Minari resolves a dataset by id, so the two can sit in one ~/.minari without either hiding the other — which matters here, because they are the input and the output of the same experiment.

Companion repos

• Policy bundles for eight MuJoCo environments: ccnets/causal-gpt-rl
• Unity ML-Agents trajectories, with a calibrated quality ladder: ccnets/causal-gpt-rl-unity-datasets
• Code: ccnets-team/causal-gpt-rl

한국어 카드 / Korean edition: README.ko.md