README.md

metadata

title: Tokescope (full)
emoji: 🔬
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 5.50.0
app_file: gradio_app.py
pinned: false
license: mit
short_description: Interactive observability for transformer internals

Tokescope — full build (private)

Interactive observability for transformer internals: pick a focus word and a control, and watch where the model ranks each — layer by layer — as it writes, with rank‑band heatmaps, a per‑layer trace, a latency/closest‑cells log, and threshold suppress / stop interventions. Runs ungated Qwen/Qwen2.5-0.5B-Instruct on CPU by default; set the GRADIOGRAPHER_MODEL variable to swap in any compatible decoder (e.g. a gemma-3 model, with access).

This private/gated build carries both projection methods side by side:

• lm — the standard logit‑lens (final_norm + lm_head).
• interp — an experimental, lm_head‑free projection. Held out of the public build (a cheap way to explore the rank neighborhood of arbitrary tokens is a mild dual‑use surface); kept here, behind access control, for research.

Where the two disagree is the interesting part.

---

Built by James J. Davison, with Claude (Anthropic) as coding collaborator. Responsibility for the code and its design choices rests with the human author. MIT licensed. © 2026 James J. Davison.