Skip to content

Quickstart

Install

pip install reptimeline

Command-Line Usage

# Track specific concepts across training checkpoints
python -m reptimeline \
  --checkpoint-dir checkpoints/ \
  --concepts king queen love hate

# Full analysis with all 63 primitives, overlay, and plots
python -m reptimeline \
  --checkpoint-dir checkpoints/ \
  --primitives --overlay --plot \
  --max-checkpoints 8 \
  --output timeline.json

Python API

from reptimeline import TimelineTracker, TriadicExtractor

# 1. Extract representations from checkpoints
extractor = TriadicExtractor()
concepts = ["king", "queen", "love", "hate", "dog", "cat"]
snapshots = extractor.extract_sequence("checkpoints/", concepts)

# 2. Analyze evolution
tracker = TimelineTracker(extractor)
timeline = tracker.analyze(snapshots)

# 3. Inspect results
timeline.print_summary()
# Births: 2632, Deaths: 1732, Connections: 1378, Phase transitions: 3

Visualization

from reptimeline.viz import plot_phase_dashboard, plot_churn_heatmap

plot_phase_dashboard(timeline, save_path="phase.png")
plot_churn_heatmap(timeline, save_path="churn.png")

See Visualization for all 4 plot types.

Bottom-Up Discovery

No labels needed -- discover what the model learned:

from reptimeline.discovery import BitDiscovery

discovery = BitDiscovery()
report = discovery.discover(snapshots[-1], timeline=timeline)
discovery.print_report(report)
# Discovers: bit semantics, dual pairs, dependencies, 3-way interactions

See Discovery Pipeline for the full pipeline.

Next Steps