📁Data Locations

Understanding where SLM Lab saves data helps you analyze results, reproduce experiments, and share findings.

Output Directory Structure

After training, results are saved to data/{spec_name}_{timestamp}/:

data/ppo_cartpole_2026_01_30_221924/
├── ppo_cartpole_spec.json                              # Original spec
├── ppo_cartpole_t0_spec.json                           # Trial spec (for reproduction)
├── ppo_cartpole_t0_trial_graph_mean_returns_vs_frames.png    # Trial training curve
├── ppo_cartpole_t0_trial_graph_mean_returns_ma_vs_frames.png # Trial moving average
├── ppo_cartpole_t0_trial_metrics_scalar.json           # Trial metrics
│
├── graph/                  # Per-session graphs
│   └── ppo_cartpole_t0_s0_session_graph_*.png
│
├── info/                   # Session data and experiment results
│   ├── ppo_cartpole_t0_s0_session_df_train.csv        # Training metrics
│   ├── ppo_cartpole_t0_s0_session_df_eval.csv         # Evaluation metrics
│   └── experiment_df.csv                               # Search results (if search mode)
│
├── log/                    # TensorBoard event files
│   └── events.out.tfevents.*
│
└── model/                  # PyTorch checkpoints
    ├── ppo_cartpole_t0_s0_ckpt-best_net_model.pt      # Best model (session 0)
    └── ppo_cartpole_t0_s0_net_model.pt                # Final model (session 0)

Naming convention: {spec_name}_t{trial}_s{session}_{type}.{ext}

• t0 = Trial index 0

• s0 = Session index 0

Key Files Explained

Session DataFrames (*_session_df_{train,eval}.csv)

Two variants are saved per session:

• *_session_df_train.csv: Metrics during training rollouts

• *_session_df_eval.csv: Metrics during evaluation (more stable)

Time series of metrics, one row per checkpoint:

Column	Description
frame	Total frames processed
total_reward	Episode reward at this checkpoint
total_reward_ma	100-checkpoint moving average of reward
loss	Training loss
fps	Frames per second
wall_t	Wall-clock time (seconds)
t	Step counter
epi	Episode count
opt_step	Optimization steps
lr	Current learning rate
entropy	Policy entropy (policy gradient algorithms)
clip_eps	PPO clipping epsilon (PPO only)

Note: Some columns are algorithm-specific (e.g., clip_eps only appears for PPO, entropy for policy gradient algorithms).

Example usage:

import pandas as pd

df = pd.read_csv('data/ppo_cartpole_2026_01_30_221924/info/ppo_cartpole_t0_s0_session_df_eval.csv')
print(f"Final reward MA: {df['total_reward_ma'].iloc[-1]:.2f}")
print(f"Training time: {df['wall_t'].iloc[-1] / 60:.1f} minutes")

Trial Metrics (*_trial_metrics_scalar.json)

Summary statistics aggregated across sessions:

{
  "strength": 245.3,
  "efficiency": 0.82,
  "stability": 0.95,
  "consistency": 0.91,
  "final_return_ma": 267.5
}

See Performance Metrics for detailed definitions.

Experiment DataFrame (experiment_df.csv)

For search mode, contains all trial results sorted by performance:

Column	Description
trial	Trial index
{search_var}	Hyperparameter values tried
total_reward_ma	Final performance
strength, efficiency, etc.	Performance metrics

Example usage:

import pandas as pd

df = pd.read_csv('data/ppo_breakout_search_2026_01_30/info/experiment_df.csv')
best_trial = df.iloc[0]  # Sorted best-first
print(f"Best trial: {best_trial['trial']}")
print(f"Best lam: {best_trial['agent.algorithm.lam']}")
print(f"Best reward: {best_trial['total_reward_ma']:.2f}")

Model Checkpoints

Two checkpoints are saved per session:

• *_ckpt-best_net_model.pt: Model with highest total_reward_ma during training

• *_net_model.pt: Model at the end of training (no ckpt- prefix)

Usually these are similar, but "best" is useful if performance dropped near the end of training (policy collapse).

Loading a checkpoint:

from slm_lab.agent.net import net_util

# Load weights into a network
net_util.load(agent.algorithm.net, 'data/ppo_cartpole_2026_01_30_221924/model/ppo_cartpole_t0_s0_ckpt-best_net_model.pt')

# Or load all nets for an algorithm (recommended)
net_util.load_algorithm(agent.algorithm)

Spec File (*_spec.json)

The complete spec used for this run, including:

• All hyperparameters

• Git SHA of the code version

• Random seeds

This file is all you need to reproduce the experiment.

Quick Reference

What you want	Where to find it
Final reward	total_reward_ma in last row of info/*_session_df_eval.csv
Training curves	*_trial_graph_mean_returns_ma_vs_frames.png (root folder)
Best hyperparameters	First row of info/experiment_df.csv
Model for inference	model/*_ckpt-best_net_model.pt
Reproduce this run	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
TensorBoard data	log/ folder

Working with Results

Using SLM Lab's Analysis Module

SLM Lab uses Plotly for visualization. You can regenerate graphs using the retro_analysis module:

uv run python -c 'from slm_lab.experiment import retro_analysis; retro_analysis.retro_analyze("data/ppo_cartpole_2026_01_30_221924")'

See Post-Hoc Analysis for more details.

Viewing Graphs

The generated PNG graphs can be opened directly. For interactive viewing, SLM Lab also generates HTML files with Plotly's interactive features (zoom, hover data).

Export for Publication

SLM Lab automatically creates a zip archive of the data folder. For sharing results:

1. Upload the zip to cloud storage or HuggingFace

2. Share the spec file for reproducibility

3. Reference the git SHA for exact code version

See Public Benchmark Data for examples.