SRT-Adapter v1.0 — Live Demo

Interactive demo for the Semiotic-Reflexive Transformer Adapter (v1.0 = v15a checkpoint) bolted onto a frozen Qwen/Qwen2.5-7B.

Three demos in one Space:

1. Per-token readouts (preserved from v8a) — paste a passage, get
   • per-token $\hat{r}$ reflexivity heatmap
   • per-token $P(\text{supercritical})$ regime probability
   • layer-wise divergence norms (7 / 14 / 21)
   • 64-D community-vector summary
   • JSON / Markdown export of the trace
2. Sentence similarity — enter 2–32 sentences, get the pairwise cosine grid of community_output.encoded mean-pooled embeddings (the head v1.0's contrastive training optimizes; English-mean Spearman ≈ 0.589 across 17 MTEB STS splits).
3. Community clusters (live Cap-A) — KMeans on the same embeddings. Default sample is Banking77 (intent classification with NMI / ARI vs gold labels, NMI ≈ 0.539 ± 0.008 over 3 seeds at 20 intents × 2 sentences); switch to Your text to cluster your own one-line-per-passage input.

This Space targets ZeroGPU. First request after a cold start downloads Qwen2.5-7B (~15 GB) into the Space cache, which can take a few minutes; subsequent requests return in under a second.

For the paper, weights, and reproducible benchmarks, see the model card.