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=8 (ORACLE-DRIFT)
Minimal-LoRA upper-bound experiment: rank 8, α=32 (α/r=4) on
Qwen3-8B q_proj/v_proj across all 36 transformer blocks (~3.83M
LoRA parameters).
Outcome branch: ORACLE-DRIFT. Probe-G-oracle re-confirm drifts from Stage 1a's 0.7871 to 0.7277 (-5.94pp), exceeding the pre-registered ±2pp tolerance. LoRA at this scale alters base text reading on Qwen3-8B; consumer Probe-G$_{\mathrm{neutral}}$ falls to 0.4827 (-0.029pp from R0).
This is the negative checkpoint demonstrating that going beyond the minimal rank-4 setting amplifies the LLM-perturbation trade-off rather than improving consumer-side audio use. The result is consistent with concurrent LoRA-rank trade-off findings (Rathore et al. 2025, arXiv:2512.15634).
Files: PEFT LoRA weights.
| Value | |
|---|---|
| Oracle re-confirm | 0.7277 (-5.94pp drift) |
| Probe-G$_{\mathrm{neutral}}$ | 0.4827 |
| Probe-G$_{\mathrm{total}}$ | 0.5938 |
| Probe-G$_{\mathrm{explicit}}$ | 0.7050 |
Code / paper: https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress
License: CC-BY-NC-4.0.
- Downloads last month
- 30
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support