Modelpublic

BLANK/mario-dqn-workshop

sourceHugging Faceotherupdated 24d agoView on Hugging Face

0likes18downloads

Model Card

mario-dqn-workshop

Educational Super Mario Bros. World 1-1 Deep Q-Network (DQN) checkpoints for the SLM-RL Colab workshop.

These weights exist so a 15–20 minute live-training cell can start from a warm-start policy and so a final evaluation cell can show a stronger public policy. They are not a product, a Nintendo-approved model, or a replacement for the workshop’s Atari RAM-vector teacher.

Files

File	Role
`warm-start.chkpt`	Early World 1-1 play. Default starting point for live training.
`final.chkpt`	Later World 1-1 play. Default public evaluation policy.
`config.json`	Action map, frame stack, and preprocessing.
`checksums.json`	SHA-256 pins for the checkpoint files.
`metrics.json`	Recorded evaluation distances and rewards.

A random / untrained baseline is generated locally. It is not stored here.

Environment

—Title: Super Mario Bros. World 1-1 (SuperMarioBros-1-1-v0)
—Observation: 4 stacked 84×84 grayscale frames
—Actions: RIGHT, RIGHT+A
—Encoder: Nature-style CNN
—Frame skip: 4

How the notebook uses this repo

The Colab cell MARIO_MODEL_REPO defaults to BLANK/mario-dqn-workshop and pins MARIO_MODEL_REVISION. Downloads are anonymous. A checksum mismatch or network failure falls back to the committed workshop clips.

Live training writes local-trained.chkpt on the Colab disk. Participant uploads stay off unless the participant supplies their own token.

Limitations

—A short workshop run does not equal full training.
—Super Mario Bros is owned by Nintendo. These weights are educational evidence only.
—The workshop’s production teacher remains the Atari RAM-vector MLP in slm_rl.teachers.dqn.