MakazhanAlpamys/Soup v0.23.0

v0.23.0 — Advanced Quantization + Data Pipeline

on GitHub

What's New

AWQ/GPTQ Export

• soup export --format awq — AWQ quantization via autoawq (pip install 'soup-cli[awq]')
• soup export --format gptq — GPTQ quantization via auto-gptq (pip install 'soup-cli[gptq]')
• New flags: --bits (4/8), --group-size, --calibration-data, --calibration-samples
• Calibration data path traversal protection

Sample Packing

• packing: true in training config — pack multiple short samples into one sequence
• Works with SFT and Pretrain tasks via TRL's native packing support
• Warning emitted if max_length < 256

Data Split

• soup data split --val 10 --test 10 — split datasets into train/val/test
• Random and stratified splitting strategies
• Reproducible with --seed, supports absolute counts with --absolute

Curriculum Learning

• curriculum: true — sort dataset by difficulty (easy → hard) for better convergence
• Metrics: length (default), perplexity, loss
• Configurable bucket count for staged training

Stats

• 1970 tests across 70 test files
• All CI green (Python 3.9/3.11/3.12 × Ubuntu/Windows/macOS)

Install / Upgrade

pip install --upgrade soup-cli
# or
pip install soup-cli==0.23.0