MakazhanAlpamys/Soup v0.12.0

v0.12.0: Alignment Methods + Advanced PEFT

on GitHub

What's New

v0.11.0 — Alignment Methods

• ORPO (task: orpo) — odds ratio preference optimization, no reference model needed
• SimPO (task: simpo) — simple preference optimization with length-normalized rewards
• IPO (task: ipo) — identity preference optimization with stronger regularization
• Templates: soup init --template orpo/simpo/ipo
• Init wizard, train routing, and sweep shortcuts for all three

v0.12.0 — Advanced PEFT

• DoRA (use_dora: true) — weight-decomposed LoRA across all 9 trainers
• LoRA+ (loraplus_lr_ratio) — different learning rates for A and B matrices
• GaLore (use_galore: true) — memory-efficient full-parameter training on consumer GPUs

Security

• experiment_name path traversal validation (blocks /, \, :, null bytes)
• GaLore optim_args type enforcement before string interpolation

Stats

• 877 tests across 42 test files (was 746 / 38)
• 56.98% coverage (above 50% threshold)
• All 5 reviews passed: python-review, code-review, security-review, tdd, verify

Install / Upgrade

pip install --upgrade soup-cli

Full Changelog

v0.10.10...v0.12.0