🔬Post-Hoc Analysis

After training, SLM Lab automatically generates graphs and metrics. You can regenerate these anytime with updated styling or new metrics.

What SLM Lab Generates

Every training run produces:

Training curves showing reward over time:

Example Training Curve

Moving average curves for smoother visualization:

Example Moving Average

Regenerating Analysis

Use retro_analyze to regenerate graphs and metrics without re-running training:

Basic Usage

uv run python -c 'from slm_lab.experiment import retro_analysis; retro_analysis.retro_analyze("data/ppo_lunar_2026_01_30_221924")'

This regenerates:

• All graphs (PNG and HTML)

• Trial-level aggregated metrics

• Experiment summary (for search runs)

What Gets Regenerated

Artifact	Location	Effect
Trial graphs	*_trial_graph_*.png (root)	Overwritten
Trial metrics	*_trial_metrics_scalar.json (root)	Overwritten
Session graphs	graph/*_session_graph_*.png	Overwritten
Session data	info/*_session_df_{train,eval}.csv	Preserved
Model checkpoints	model/*.pt	Preserved
Spec file	*_spec.json (root)	Preserved

Safe to run: Retro analysis only overwrites derived data. Your raw session data, trained models, and spec files are never modified.

Common Use Cases

Update Graph Styling

SLM Lab updates Plotly styling periodically. Regenerate graphs to get the latest look:

# Pull latest code
git pull
uv sync

# Regenerate all experiments in data/
for dir in data/*/; do
    uv run python -c "from slm_lab.experiment import retro_analysis; retro_analysis.retro_analyze('$dir')"
done

Recompute Metrics After Code Changes

If you modify the analysis module (e.g., add a new metric):

# After modifying slm_lab/experiment/analysis.py
uv run python -c 'from slm_lab.experiment import retro_analysis; retro_analysis.retro_analyze("data/ppo_lunar_2026_01_30_221924")'

# Check updated metrics
cat data/ppo_lunar_2026_01_30_221924/ppo_lunar_t0_trial_metrics_scalar.json

Generate Publication-Quality Graphs

For papers or presentations, you may want higher-resolution or different formats:

from slm_lab.experiment import retro_analysis

# Regenerate with custom settings
retro_analysis.retro_analyze('data/ppo_lunar_2026_01_30_221924')

# The HTML files support interactive exploration
# PNG files are ready for documents

Batch Processing

Process multiple experiments:

import os
from slm_lab.experiment import retro_analysis

data_dirs = [d for d in os.listdir('data') if os.path.isdir(f'data/{d}')]

for data_dir in data_dirs:
    print(f"Processing {data_dir}...")
    retro_analysis.retro_analyze(f'data/{data_dir}')

Troubleshooting

"Session data not found"

Ensure the data folder contains *_session_df_train.csv and *_session_df_eval.csv files in the info/ subfolder. These are required for analysis.

Graphs look wrong

Check that your SLM Lab version matches the data format. Very old experiments may need manual migration.