LensLapse

An in-browser logit lens over training checkpoints: scrub a slider across a model’s public checkpoints and watch per-layer, per-position predictions crystallize from frequency bias into knowledge — plus a pip install lenslapse local probe server for your own models.

• Paper (AACL-IJCNLP 2026 System Demonstrations): https://aclanthology.org/2026.aacl-demo.9/

• Live demo: https://lenslapse.iamtatsuki05.com/

• Project page: https://lenslapse.iamtatsuki05.com/about/

• Source: https://github.com/iamtatsuki05/lenslapse

• PyPI: https://pypi.org/project/lenslapse/

Guides

• Advanced usage
  • Develop
  • Convert checkpoints & precompute
  • Add your own model (Hub or local)
  • Heavy models: the local probe server
  • Script it: the CLI mirrors the UI
  • Reproducibility of probes
  • Benchmark
  • Building the docs
  • Publishing to PyPI
• Zero-cost deployment
  • 1. Upload converted checkpoints to Hugging Face Hub
  • 2. Point the app at the repo
  • 3. Publish the site on GitHub Pages
  • Verification checklist after deploy
• LensLapse ONNX checkpoints — Pythia suite across training, plus single-checkpoint architecture coverage
  • Contents
  • Fidelity
  • Reproduce
  • License and attribution

API reference

• lenslapse
  • lenslapse package