Reinforcement Learning
English
steve1
minecraft
model-zoo
interpretability

MultiSTEVE-1s Model Zoo

This repository contains the MultiSTEVE-1s model zoo: 140 STEVE-1/VPT Minecraft agents trained with fixed hyperparameters and controlled randomness. The release is intended for reproducibility, interpretability, and controlled comparisons of training nondeterminism.

Contents

  • 01_init_seed_only: dataset seed fixed, weight-initialisation seed varies.
  • 02_data_seed_only: weight-initialisation seed fixed, dataset seed varies.
  • 03_init_and_data: both seeds vary together.
  • 04_single_lsb_flip: seeds fixed, with one selected MineCLIP embedding-layer weight bit-flipped before training.
  • 05_final_init_and_data_100: 100 additional both-seed runs for outlier search.

Runs contain checkpoints named STEVE-1_snapshot_<frames>.weights, data-order logs, and training logs. Snapshot 000000000 is the initial checkpoint before training; the main final checkpoint is 160005120.

Usage

Use the supplementary code release and place this repository under:

data/model_zoo/

The expected layout and evaluation commands are documented in the code release files:

  • README.md
  • docs/data_and_weights.md
  • docs/training.md
  • docs/evaluations.md
  • docs/reproducibility.md

Limitations

The agents act only in MineRL/Minecraft and are not general-purpose policies. Bitwise reproduction is sensitive to CUDA, GPU, driver, PyTorch, and MineRL versions. The checkpoints are research artefacts for studying model variation and interpretability, not safety-certified systems.

Licence And Terms

These checkpoints are research artefacts derived from VPT, STEVE-1, MineCLIP/MineDojo, MineRL, OpenAI contractor-demonstration workflows, and Minecraft-related assets. This repository uses license: other under the MultiSTEVE-1 research artefact terms included in the repository LICENSE file.

The new MultiSTEVE-specific source-code additions are released under MIT in the supplementary code release. This model repository does not grant rights to upstream STEVE-1, VPT, MineCLIP/MineDojo, MineRL, OpenAI contractor-demonstration, or Minecraft assets beyond their own licences and terms. It is not an official Minecraft product and is not approved by or associated with Mojang or Microsoft.

Downloads last month
5
Video Preview
loading

Dataset used to train randomhuggingfaceuser1273823147/MultiSTEVE-1s