feat[cv2-7]: Scale up distillation data (5500 train utterances, 2-GPU sharded training + checkpoint averaging)
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 0
- Forks
- 0
- PR merge metrics
- No merged PRs in 30d
Description
Goal
Iteration 7 of SPADE-on-CosyVoice2: scale the distillation training data from 1500 to 5500 LibriSpeech train-clean-100 utterances. Train two shards in parallel on 2x L4 (7 epochs each, different seeds), average the checkpoints, and evaluate WER/RTF on 200 dev-clean utterances.
Acceptance
- data_prep supports multi-parquet train sources + separate eval source + train sharding
- Two 7-epoch distillation runs (shard0/shard1) complete on 2 GPUs
- Checkpoint averaging produces the final model
- Eval WER improves vs the 1500-utterance result (0.413) with teacher baseline 0.316
Notes
Iteration 7 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 at the existing data_prep entry point and trace how the 1500-utterance train/eval workflow is configured. Verify the train sources, sharding, two 7-epoch runs, checkpoint averaging, and WER/RTF evaluation against the stated baselines. Done means the 5500-utterance workflow completes and improves WER over 0.413.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100