Reinforcement Learning
stable-baselines3
ppo
stable-retro
rlab
super-mario-bros
nes
SuperMarioBros-Nes-v0
Instructions to use tsilva/SuperMarioBros-Nes-v0_Level1-2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- stable-baselines3
How to use tsilva/SuperMarioBros-Nes-v0_Level1-2 with stable-baselines3:
from huggingface_sb3 import load_from_hub checkpoint = load_from_hub( repo_id="tsilva/SuperMarioBros-Nes-v0_Level1-2", filename="{MODEL FILENAME}.zip", ) - Notebooks
- Google Colab
- Kaggle
SuperMarioBros-NES Level1-2 PPO
PPO policy checkpoint for completing SuperMarioBros-Nes-v0 Level1-2 with supermariobrosnes-turbo, trained with rlab.
Quick Start
Install rlab once, import the ROM, then play or evaluate this checkpoint directly from Hugging Face:
uv tool install --from git+https://github.com/tsilva/rlab rlab
rlab import-roms ~/roms --game SuperMarioBros-Nes-v0
rlab play hf://tsilva/SuperMarioBros-Nes-v0_Level1-2
rlab eval hf://tsilva/SuperMarioBros-Nes-v0_Level1-2
Evaluation Results
eval_profile |
episodes |
seed_start |
completion_rate |
max_x_max |
reward_mean |
checkpoint_step |
|---|---|---|---|---|---|---|
local release eval, deterministic policy |
100 | 10000 | 100/100 episodes |
3129 | 3148.550 | 7500000 |
This release uses the step-7500000 checkpoint from the W&B run below. The training run reached the Level1-2 early-stop gate at 7,506,208 timesteps; this staged release eval was run locally because the run's post-training checkpoint eval crashed after its first checkpoint during W&B summary bookkeeping.
Environment Details
| Setting | Value |
|---|---|
env_provider |
supermariobrosnes-turbo |
env_id |
supermariobrosnes-turbo:SuperMarioBros-Nes-v0 |
game |
SuperMarioBros-Nes-v0 |
state |
Level1-2 |
preprocessing |
crop top 32 px, grayscale, resize to 84 x 84 |
frame_stack |
4 |
frame_skip |
4 |
max_pool_frames |
disabled |
policy_observation_layout |
channel-first (4, 84, 84) |
action_set |
simple |
reward_mode |
score |
reward_shaping |
score-progress reward, terminal_reward=50, death_penalty=25, reward_scale=10 |
max_episode_steps |
4500 |
train_done_on_events |
life loss, level change |
eval_done_on_events |
level change |
Training Recipe
| Setting | Value |
|---|---|
goal_id |
Level1-2 |
recipe_id |
base |
checkpoint_step |
7500000 |
frame_skip |
4 |
max_episode_steps |
4500 |
reward_mode |
score |
reward_shaping |
score-progress reward, terminal_reward=50, death_penalty=25, reward_scale=10 |
done_on_events |
life loss, level change |
Provenance
| Item | Value |
|---|---|
source_project |
rlab |
goal_id |
Level1-2 |
goal_title |
Sample-efficient SuperMarioBros-NES Level1-2 completion |
wandb_run |
Level1-2_base_s1_20260704T072752Z |
wandb_artifact |
tsilva/SuperMarioBros-Nes-v0/Level1-2_base_s1_20260704T072752Z-checkpoint:step-7500000 |
eval_source |
local release eval |
- Downloads last month
- 25