feat[cv2-4]: WER-based layer importance (WLI) on the CosyVoice2 LLM
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 0
- Forks
- 0
- PR merge metrics
- No merged PRs in 30d
Description
Goal
Iteration 4 of SPADE-on-CosyVoice2: compute leave-one-out WER-based layer importance over the 24 Qwen2 LLM layers on a real-speech eval subset (synthesis + Whisper transcription), producing the WLI report that drives pruning.
Acceptance
-
spade_cosyvoice2/wli.pybypasses each LLM layer (zeroing via spade.adapters.hf.bypass_block), synthesizes eval utterances zero-shot, transcribes with Whisper, and reports per-layer WER - WLI report JSON written next to the eval list
- Layers restored after each leave-one-out run
Notes
Iteration 4 for: SPADE applied to CosyVoice 2 (arXiv:2509.20802)
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with spade_cosyvoice2/wli.py and the spade.adapters.hf.bypass_block entry point; trace the 24-layer Qwen2 LLM path and the real-speech evaluation list. Run synthesis and Whisper transcription for each leave-one-out layer, verify layers are restored after every run, and confirm the per-layer WER report is written next to the evaluation list.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 50/100