Spaces:
Running on Zero
Running on Zero
File size: 2,220 Bytes
5603210 cec8e12 5603210 ae4c8a2 5603210 ae4c8a2 cec8e12 ae4c8a2 ec32e23 5603210 cec8e12 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | ---
title: SRT Showcase
emoji: 🧭
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
python_version: '3.10'
pinned: true
hardware: zerogpu
short_description: Watch a frozen Qwen-2.5-7B think — live SRT introspection
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-showcase/resolve/main/thumbnail.png
---
# SRT Showcase — watch a frozen model think
Live, token-by-token introspection of **Qwen-2.5-7B + the SRT adapter**.
As the model generates, every token is tinted by its predictive **entropy** (the
validated uncertainty signal). At the highest-effort token positions, chosen by
an adaptive-density scheduler, the **Activation Verbalizer** decodes the model's
internal hidden state into natural language, and each verbalization carries a
**round-trip fidelity badge**: it is re-encoded and compared back to the original
hidden state, so the "this is what the model was thinking" claim is visibly
self-validating.
Features:
- Live token stream tinted by entropy or SRT divergence, with per-token hover
rollovers (entropy, divergence, reflexivity `r̂`, regime).
- Running entropy meter and entropy / divergence charts.
- Expand/collapse verbalization cards with round-trip fidelity badges.
- A/B panel: the same prompt with SRT injection on vs off (bare backbone),
seeded identically.
- A curated example gallery covering confident recall, false premises,
misconceptions, reasoning pivots, genuine uncertainty, and safety boundaries.
## Honest scope
Entropy is the load-bearing uncertainty signal. The SRT side-channels
(divergence, `r̂`, regime) and the verbalizations are shown as **observational
readouts** of internal state. This is a window into the model, not a validated
hallucination detector.
## Notes
- First request is cold (~60–90 s) while ZeroGPU acquires a GPU and the ~16 GB
backbone weights load; subsequent requests are warm.
- A second backbone copy is loaded for the Activation Verbalizer.
Source: <https://github.com/space-bacon/SRT>
|