Beyond Transcript Alignment
Collection
Frozen-frozen speech-LLM adapters + counterfactual training. Code: github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress • 8 items • Updated
BTA — Stage 6 LoRA r=4
Minimal-LoRA upper-bound experiment: rank 4, α=16 (α/r=4) on
Qwen3-8B q_proj/v_proj across all 36 transformer blocks (72 LoRA
modules total, ~1.92M LoRA parameters). The R1.8 adapter and
$C_\phi$ heads are FROZEN; only the LoRA matrices update.
Pre-registered oracle-stability gate: Probe-G-oracle re-confirm within ±2pp of the Stage 1a anchor 0.7871. Result: 0.7970 (+0.99pp drift) → PASS.
Outcome branch: weak-negative. Probe-G$_{\mathrm{neutral}}$ = 0.4926 (R0 - 0.0196), 0.4pp below the H2 NULL band on the negative side. Minimal LoRA at this rank preserves the oracle but does not recover audio-conditional consumer use of the R1.8 representation.
Files: PEFT LoRA weights (adapter_config.json, adapter_model.safetensors).
Load via PeftModel.from_pretrained on top of base Qwen/Qwen3-8B.
Files / metrics:
| Value | |
|---|---|
| Oracle re-confirm | 0.7970 |
| Probe-G$_{\mathrm{neutral}}$ | 0.4926 |
| Probe-G$_{\mathrm{total}}$ | 0.6277 |
| Probe-G$_{\mathrm{explicit}}$ | 0.7629 |
| Probe-K linear (R1.8 frozen) | 0.2265 |
| Probe-K MLP-2 (R1.8 frozen) | 0.3120 |
Code / paper: https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress
License: CC-BY-NC-4.0.
- Downloads last month
- 32
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support