rottter4585/Llasa-GRPO
🎤 Fine-tune the Llasa TTS model with GRPO using Hugging Face tools to enhance performance and evaluate rewards with Whisper ASR and WER metrics.
What's novel
🎤 Fine-tune the Llasa TTS model with GRPO using Hugging Face tools to enhance performance and evaluate rewards with Whisper ASR and WER metrics.
Code Analysis
5 files read · 2 roundsA thin fine-tuning pipeline that applies HuggingFace TRL's GRPOTrainer to the Llasa-1B TTS model, using a Whisper-based WER+NLL composite reward function to improve speech synthesis quality.
Strengths
The reward function in reward_whisper.py is the most substantive piece, combining WER and NLL into a harmonic-mean composite with tunable weights, and the overall pipeline (dataset encoding → GRPO training → inference) is coherent and functional.
Weaknesses
The project is almost entirely a wrapper around trl.GRPOTrainer with no novel algorithmic contribution; the README is heavily misleading (claims no programming needed, fake links, nonexistent features like 'Automatic Updates'), there are zero tests, and code quality issues include duplicated functions, hardcoded CUDA assumptions, and debug prints left in production code.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
16
Contributors
2
Files
8
Active weeks
6
Repository
Language
Python
Stars
2
Forks
0
License
—