---
title: SRT introspect
emoji: 🧭
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
python_version: '3.10'
pinned: true
short_description: Adaptive-density reasoning traces over a frozen Qwen-2.5-7B
hardware: zero-a10g
models:
- Qwen/Qwen2.5-7B
- RiverRider/srt-adapter-v1.0
- RiverRider/srt-nla-av-v1
tags:
- srt
- semiotic-reflexive-transformer
- interpretability
- introspection
- uncertainty
- visualization
- llm
thumbnail: https://huggingface.co/spaces/RiverRider/srt-introspect/resolve/main/thumbnail.png
---

# SRT · introspect

Live demo of the **SRT-Adapter** (Stage 3) + **Activation Verbalizer** (Stage 4)
applied to a frozen Qwen-2.5-7B backbone.

Enter a prompt → the model generates a continuation, every token is tinted by
the adapter's per-step divergence, and the scheduler picks the most novel
positions to narrate via the AV.

- **Code**: https://github.com/space-bacon/SRT
- **Adapter weights**: `RiverRider/srt-adapter-v1.0`
- **AV weights**: `RiverRider/srt-nla-av-v1`

The first request on a fresh ZeroGPU slice takes ~60–90 s (weight download).
Subsequent generations are ~7–10 s.