| --- |
| title: Induction Circuit Stability Under Fine-Tuning |
| emoji: 🔬 |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 4.37.2 |
| app_file: app.py |
| pinned: true |
| python_version: "3.10" |
| license: mit |
| short_description: Visualise induction head circuits change under fine-tuning. |
| tags: |
| - mechanistic-interpretability |
| - transformers |
| - circuits |
| - induction-heads |
| - fine-tuning |
| - safety |
| --- |
| |
|
|
| # Induction Circuit Stability Under Fine-Tuning |
|
|
| Interactive dashboard accompanying the paper *Induction Circuit Stability Under |
| Fine-Tuning: A Mechanistic Interpretability Study*. |
|
|
| ## What this does |
|
|
| Induction heads are attention heads that implement the copy-and-complete pattern: |
| if the model has seen A→B before, an induction head at the second A attends back |
| to B and predicts B will follow. This is the primary in-context learning mechanism |
| in small transformers (Olsson et al., 2022). |
|
|
| This dashboard lets you inspect: |
| - **Attention heatmap** — which tokens each head attends to on your input text. |
| - **Induction score grid** — per-head induction strength (0 = none, 1 = perfect). |
| - **Circuit diagram** — which heads are causally verified circuit members. |
|
|
| Toggle between **Pre-fine-tuning** and **Post-fine-tuning** states to see how the |
| circuit changes after training on Python code. |
|
|
| ## Load time |
|
|
| The dashboard loads ONNX models on first run (pre-cached in this Space). |
| Typical cold-start: **< 3 seconds** on Hugging Face CPU Basic hardware. |
| Inference per input: **< 500 ms**. |
|
|
| ## Links |
|
|
| - GitHub: [Mattral/mech-interp-induction](https://github.com/Mattral/Mechanistic-Interpretability-Study-Induction-Circuit-Stability-Under-Fine-Tuning) |
| - Paper: [ArXiv link TBD] |
|
|
