CV-ICRL training histories

Training histories for Context Value Informed In-Context Reinforcement Learning, by Wenhao Zhang, Shao Zhang, Xihuai Wang, Yang Li, and Ying Wen. The companion implementation is Bluixe/CV-ICRL.

This release contains six MiniGrid CV training files and three Procgen CV training files, totaling 117,485,557,026 bytes (117.49 GB / 109.42 GiB). Files preserve their original serialization and array contents. The repository contains training histories only; validation histories, model checkpoints, AD datasets, and other relabeling variants are not included. The companion training examples that use a validation file still require that separate input.

Publication status: complete (2026-10-12). All nine training files are available, and their byte sizes have been checked against the manifest. The fixed release is tagged camera-ready-v1.

Files

Family	Task / game	File	Bytes	Stored array length
MiniGrid	MiniGrid-SimpleCrossingS9N3-v0	train_traj-relabel.pkl	11,152,000,471	170,000 × 400
MiniGrid	MiniGrid-SimpleCrossingS11N5-v0	train_traj-relabel.pkl	11,152,000,471	170,000 × 400
MiniGrid	MiniGrid-BlockedUnlockPickup-v0	train_traj-relabel.pkl	11,152,000,471	170,000 × 400
MiniGrid	MiniGrid-Unlock-v0	train_traj-relabel.pkl	11,152,000,471	170,000 × 400
MiniGrid	MiniGrid-LavaCrossingS9N3-v0	train_traj-relabel.pkl	11,152,000,471	170,000 × 400
MiniGrid	MiniGrid-LavaCrossingS9N2-v0	train_traj-relabel.pkl	11,152,000,471	170,000 × 400
Procgen	bigfish	train_traj-relabel-new.npz	10,459,251,400	425 × 2,000
Procgen	starpilot	train_traj-relabel-new.npz	16,144,161,400	656 × 2,000
Procgen	miner	train_traj-relabel-new.npz	23,970,141,400	974 × 2,000

Paths are MiniGrid/<task>/train_traj-relabel.pkl and Procgen/<game>/train_traj-relabel-new.npz. Downloading to datasets/ gives the layout expected by the companion code. Full paths, byte sizes, shapes, and dtypes are recorded in manifest.json.

Array format and label semantics

Each MiniGrid pickle contains a dictionary of NumPy arrays:

Key	Dtype	Shape	Meaning
observations	uint8	[170000,400,3,7,7]	Channel-first MiniGrid observations
actions	int64	[170000,400]	Collected action indices
rewards	float64	[170000,400]	Evaluated return of the source PPO checkpoint, used as the CV training label and context scalar
dones	bool	[170000,400]	Episode termination markers

MiniGrid's rewards field has been relabeled: it is not the original environment reward. There is no separate values field. A stored 400-step row is a history segment, and need not correspond to one complete episode.

Each Procgen NPZ contains the following arrays, with N = 425 (Bigfish), 656 (Starpilot), or 974 (Miner):

Key	Dtype	Shape	Meaning
observations	uint8	[N,2000,3,64,64]	Channel-first RGB observations
actions	int64	[N,2000]	Collected action indices
rewards	float32	[N,2000]	Original environment rewards
dones	bool	[N,2000]	Episode termination markers
values	float32	[N,2000]	Evaluated source-checkpoint returns used as CV labels

Procgen's 2,000-step length is the stored stream length. The released training implementation samples H=400 windows from these streams. The values labels are checkpoint-return proxies, not independently evaluated state-conditioned values. In each game's file all rows share the same checkpoint-label sequence; Bigfish and Starpilot have 40 distinct labels, and Miner has 30. These files are the later single-game training batch selected by the authors for this release.

Download

Install the official Hub client, then download this repository's small download_data.py and manifest.json into the same directory:

pip install 'huggingface_hub>=0.36'
hf download bluixe/CV-ICRL --repo-type dataset --revision camera-ready-v1 download_data.py manifest.json --local-dir cv-icrl-data-tools
python cv-icrl-data-tools/download_data.py --dry-run
python cv-icrl-data-tools/download_data.py --family minigrid --task MiniGrid-Unlock-v0 --output-dir datasets
python cv-icrl-data-tools/download_data.py --family procgen --output-dir datasets

The downloader defaults to camera-ready-v1, resolves one immutable commit before fetching files, checks that every selected file is available with the expected byte length, and verifies the downloaded lengths. Downloads resume through the official Hub client. It does not deserialize the large arrays. The manifest also records the Hub's SHA-256 file identifiers. For all nine files, allow at least 118 GB of available space, plus filesystem/cache overhead. Training loaders can require substantial RAM because these files hold full observation arrays.

MiniGrid uses Python pickle; load it only from a trusted source. Procgen arrays can be loaded using numpy.load(path, allow_pickle=False). Loading the complete observations member allocates a large array; inspecting the ZIP/NPY headers does not require loading it.

Scope and licensing

These histories were generated through simulated MiniGrid and Procgen interaction. This release covers the CV training inputs for the six MiniGrid task types and three Procgen games above. It does not by itself include every input required to reproduce all experiments in the paper.

The released training histories are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). When reusing the histories, credit the authors of Context Value Informed In-Context Reinforcement Learning, link to this dataset and its license, and indicate any changes. See LICENSE.md. MiniGrid and Procgen are separate third-party projects; their software licenses remain with the respective projects.