🔄Resume and Replay

Resume interrupted training or replay trained models.

Resume Training

Use train@{predir} to resume from a previous run:

# Resume from latest run of this spec
slm-lab run slm_lab/spec/benchmark/ppo/ppo_cartpole.json ppo_cartpole train@latest

# Resume from specific folder
slm-lab run slm_lab/spec/benchmark/ppo/ppo_cartpole.json ppo_cartpole train@data/ppo_cartpole_2026_01_30_221924

train@latest resolves to the most recent data/{spec_name}_*/ folder.

Extending Training

To continue a completed run (e.g., 100k → 200k frames):

1. Edit the spec's max_frame

2. Resume with train@latest

Replay Mode

Use enjoy@{spec_file} to replay a trained model with rendering:

slm-lab run slm_lab/spec/benchmark/ppo/ppo_cartpole.json ppo_cartpole enjoy@data/ppo_cartpole_2026_01_30_221924/ppo_cartpole_t0_spec.json

In enjoy mode, the spec file and spec name args are ignored—everything loads from the enjoy@ path.

Enjoy mode finds the best session (by total_reward_ma) and loads its ckpt-best model checkpoint.

Replaying Published Benchmarks

Download and replay trained agents from HuggingFace:

slm-lab list              # List available experiments
slm-lab pull ppo_cartpole # Download trained model
slm-lab run slm_lab/spec/benchmark/ppo/ppo_cartpole.json ppo_cartpole enjoy@data/ppo_cartpole_2026_01_30_221924/ppo_cartpole_t0_spec.json

See Public Benchmark Data for the full list.