Sample-efficient SuperMarioBros-NES Level1-3 completion β GradLab PPO @ 9.31M env steps
An immutable PPO research release for SuperMarioBros-Nes-v0 / Level1-3, backed by accepted exact-contract evaluation evidence and a separately labeled representative replay.
Identity
| Item | Value |
|---|---|
| Goal | Level1-3 β Sample-efficient SuperMarioBros-NES Level1-3 completion |
| Environment | SuperMarioBros-Nes-v0 (env-supermariobrosnes-turbo-emu:SuperMarioBros-Nes-v0) |
| Trainer | GradLab |
| Algorithm | PPO |
| Model class | gradlab.ppo.GradLabPPO |
| Compatibility | SB3-compatible |
| Checkpoint | Exact step 9310208 |
| Immutable release | v1 |
| Release tier | research |
| Lineage | 9bfbd288fc47add5136acb6453aa15c0f379ea3e1be2bc1d0be2157bad499b4b |
| Environment index | GradLab β SuperMarioBros-Nes-v0 |
Evaluation Status
Status: Accepted under a stochastic full evaluation of 100 episodes.
| Metric | Meaning | Observed | Requirement | Outcome |
|---|---|---|---|---|
eval/success/min |
Full-eval start success rate min | 100.0% | >= 1.0 |
Pass |
Ranking metrics:
| Metric | Meaning | Observed | Direction |
|---|---|---|---|
leader/step |
Leader checkpoint step | 9310208 | min |
eval/return/mean |
Full-eval return mean | 3.37341 | max |
The normalized per-episode results, aggregates, rules, contracts, outcomes, and authoritative hashes are in evaluation_evidence.json. Evidence SHA-256: da4d2896a728aa8eeb97dc614c0b5614df422eaba44295350a041ec560e0933d.
Trainer, Algorithm, and Compatibility
- Trainer: GradLab
- Algorithm: PPO
- Model class:
gradlab.ppo.GradLabPPO - Compatibility: SB3-compatible
- Hugging Face library:
gradlab
Release History
| Release | Tier | Trainer / algorithm | Lineage | Checkpoint step | Evidence |
|---|---|---|---|---|---|
v1 |
research | GradLab PPO | 9bfbd288 |
9310208 | accepted |
Comparable Changes
Not compared: no prior comparable release in goal repository. Releases are compared only when lineage, run, seed, and evaluation-contract hash all match.
Representative Replay
Watch on YouTube. The root replay.mp4 is one validated training episode. It is representative media and is not evaluation or promotion evidence.
| Item | Value |
|---|---|
| Outcome | success |
| Goal success | True |
| Seed | 65001 |
| Start | Level1-3 |
| Policy steps | 430 |
| Episode return | 3.373570152703678 |
| Player runtime | {"asset":{"filename":"rom.nes","game":"SuperMarioBros-Nes-v0","provider_rom_identity":"facee9c577a5262dbe33ac4930bb0b58c8c037f7","provider_rom_identity_algorithm":"sha1-provider-body-v1","schema_version":2,"sha256":"f61548fdf1670cffefcc4f0b7bdcdd9eaba0c226e3b74f8666071496988248de","size_bytes":40976},"contract_mode":"training","device_type":"cpu","environment_hash":"sha256:1ac0cf02332560448534cd50b188cb4c4fb8e868341a6af0d270f5eeb7d3fcfe","execution_target":"local_player","overrides":[],"provider_id":"env-supermariobrosnes-turbo-emu","provider_version":null,"qualified_environment_id":"env-supermariobrosnes-turbo-emu:SuperMarioBros-Nes-v0","runtime_image_digest":"","runtime_versions":{"env_breakoutatari2600_turbo_native":"0.5.15","stable_baselines3":"2.9.0","stable_retro_turbo":"1.0.1.post48","supermariobrosnes_turbo":"0.7.4","vizdoom_turbo":"1.3.0.post30"},"seed":65001,"source":{"distribution":"gradlab","kind":"installed_distribution","source_tree_sha256":"c0ba50b1a6b4153ab3fed7fcbd4683612b6efa4c11ae16f7ee523e2591dcc9be","version":"0.2.2"}} |
Quick Start
Use this GradLab source revision to load the recorded recipe.
Use a legally obtained ROM that matches the game image recorded with this checkpoint.
git clone https://github.com/tsilva/gradlab.git
cd gradlab
git checkout 3a803d249661809d5c00c28a529f4752dc6b9ca3
uv sync --frozen
uv run --frozen gradlab play hf://tsilva/SuperMarioBros-Nes-v0_Level1-3@v1 --rom-path ~/roms/SuperMarioBros.nes
The immutable model reference is hf://tsilva/SuperMarioBros-Nes-v0_Level1-3@v1.
Full Evaluation
Evaluation used stochastic action selection for all 100 declared episodes. See evaluation_evidence.json for every normalized episode and all materialized acceptance and ranking outcomes.
Contracts
| Item | Value |
|---|---|
| Environment hash | sha256:b17294ed76dbceb76640d4a008a3d80bda01ea4208ccefe36c61401f154386ff |
| Preprocessing and model inputs | {"model_inputs":null,"preprocessing":{"frame_skip":4,"frame_stack":4,"max_pool_frames":false,"obs_copy":"safe_view","obs_crop":[32,0,0,0],"obs_crop_fill":0,"obs_crop_mode":"mask","obs_grayscale":true,"obs_resize":[84,84],"obs_resize_algorithm":"area","pipeline":"env_supermariobrosnes_turbo_emu_native_vec_env","policy_observation_layout":"channel_first","sticky_action_prob":0.0}} |
| Action semantics | {"contract_hash":"66617eeaf6bc3698b978d68b40d4b0dd068abf173cb739a7ed461c5a706c5a9f","execution_hash":"63f0309d266a73e07fc56f908bbcb76a0d0193d0c0b6a2e5b92723d4533c5924","policy":{"codec":{"type":"identity"},"conditional_overrides":[],"semantics":{"encoding":"explicit","entries":[{"controls":[{"atoms":[],"inputs":[],"player":1}],"label":"noop","semantic_id":"noop","value":0},{"controls":[{"atoms":["right"],"inputs":["right"],"player":1}],"label":"right","semantic_id":"right","value":1},{"controls":[{"atoms":["right","b"],"inputs":["right","b"],"player":1}],"label":"right b","semantic_id":"right_b","value":2},{"controls":[{"atoms":["right","a"],"inputs":["right","a"],"player":1}],"label":"right a","semantic_id":"right_a","value":3},{"controls":[{"atoms":["right","a","b"],"inputs":["right","a","b"],"player":1}],"label":"right a b","semantic_id":"right_a_b","value":4},{"controls":[{"atoms":["a"],"inputs":["a"],"player":1}],"label":"a","semantic_id":"a","value":5},{"controls":[{"atoms":["left"],"inputs":["left"],"player":1}],"label":"left","semantic_id":"left","value":6}],"status":"available"},"space":{"dtype":"int64","n":7,"start":0,"type":"discrete"}},"provider":{"mode":"custom_discrete","preset":"basic","provider_id":"env-supermariobrosnes-turbo-emu","semantics":{"encoding":"explicit","entries":[{"controls":[{"atoms":[],"inputs":[],"player":1}],"label":"noop","semantic_id":"noop","value":0},{"controls":[{"atoms":["right"],"inputs":["right"],"player":1}],"label":"right","semantic_id":"right","value":1},{"controls":[{"atoms":["right","b"],"inputs":["right","b"],"player":1}],"label":"right b","semantic_id":"right_b","value":2},{"controls":[{"atoms":["right","a"],"inputs":["right","a"],"player":1}],"label":"right a","semantic_id":"right_a","value":3},{"controls":[{"atoms":["right","a","b"],"inputs":["right","a","b"],"player":1}],"label":"right a b","semantic_id":"right_a_b","value":4},{"controls":[{"atoms":["a"],"inputs":["a"],"player":1}],"label":"a","semantic_id":"a","value":5},{"controls":[{"atoms":["left"],"inputs":["left"],"player":1}],"label":"left","semantic_id":"left","value":6}],"status":"available"},"space":{"dtype":"int64","n":7,"start":0,"type":"discrete"},"table_hash":"2eaa8ce13795d654097e6fbeb16460de8ae78f0af39b7f88259bc51604504134"},"requested":{"meanings":["noop","right","right_b","right_a","right_a_b","a","left"],"mode":"custom_discrete","preset":"basic","table":[[],["RIGHT"],["RIGHT","B"],["RIGHT","A"],["RIGHT","A","B"],["A"],["LEFT"]],"table_hash":"2eaa8ce13795d654097e6fbeb16460de8ae78f0af39b7f88259bc51604504134"},"schema_version":2,"semantic_hash":"ac72687dc21c167f0c05d5bd0dd7efbd858c1bccdc7903f1d78cacf1be3829da"} |
Provenance
| Item | Value |
|---|---|
| Source | gradlab |
| Run | run |
| Recipe | ppo-sil-actor-only |
| Seed | 1801 |
| Source base commit | 54215bf8306c2501577bf3b01ebca5f3135d69df |
| Source overlay SHA-256 | babe9450334e6b9cdcff31371c81668ee678c1544ad3db7dd0bbdea0b1a7f22d |
| Evaluated artifact | https://pub-fc35c0b186ce4aad8eea5e93d38c99db.r2.dev/runs/gradlab-04bbb7de64fe4f0c9e906396a59044d5/checkpoints/9310208-d43b658e042208dbbec206378fba415123375c0f9732350cfca662e9a81c9c30/model.zip |
Limitations
Evaluation establishes performance only for the published environment, policy, action-selection, start, boundary, and evaluation contracts. It does not establish generalization to other tasks or contracts. The representative replay cannot establish or change acceptance, ranking, or promotion.
Files
| File | Purpose |
|---|---|
model.zip |
Immutable policy checkpoint |
model.json |
Trainer, algorithm, model class, checkpoint identity, and provenance |
recipe.json |
Materialized training, playback, and evaluation contracts |
evaluation_evidence.json |
Normalized authoritative evaluation evidence and rule outcomes |
release_manifest.json |
Release identity, evaluation evidence, and artifact hashes |
replay.mp4 |
Browser-safe representative completed episode |
LICENSE |
License for GradLab-authored weights and publication material |
Licensing
GradLab-authored policy weights and publication material are licensed under the MIT License in LICENSE. Emulator/runtime software and game assets remain governed by their own licenses and terms. This repository does not redistribute a game ROM.