What's New
Data Quality Filters
soup data filter --coherence 0.3— filter low-quality rows by coherence scoresoup data filter --perplexity 500— filter by perplexity (uses GPT-2 by default)--score-onlyflag adds quality scores without removing rows
Audio Modality
modality: audiofor Qwen2-Audio, Whisper fine-tuning- New
audiodata format +soup init --template audio pip install 'soup-cli[audio]'
SGLang Backend
soup serve --backend sglang— high-throughput inference- Tensor parallelism, OpenAI-compatible API
pip install 'soup-cli[sglang]'
Local Server Data Generation
soup data generate --provider server— use Ollama, soup serve, LM Studio, etc.- No API key required for localhost
Security
- Server provider SSRF validation
- Audio path traversal protection
- trust_remote_code warnings for audio/SGLang
- SGLang model path SSRF validation
Install / Upgrade
pip install --upgrade soup-cli1369 tests | 56 test files | 58.8% coverage