MakazhanAlpamys/Soup v0.24.0

v0.24.0 — Dataset Hub, Freeze Training, Loss Watchdog

on GitHub

What's New

HuggingFace Dataset Hub Browser

• soup data search — search HF Hub for datasets (sort by downloads/likes/trending)
• soup data preview — view remote dataset metadata, splits, features
• soup data download — stream HF datasets to local JSONL (with optional format conversion)

Freeze Training

• freeze_layers: 24 — freeze first N layers, train only the rest
• freeze_ratio: 0.75 — freeze 75% of layers from the bottom
• Closes the gap with LLaMA-Factory's finetuning_type: freeze

Loss Watchdog

• loss_watchdog: true — auto-stop training on loss spikes
• Configurable threshold and patience
• Closes the gap with Axolotl's loss_watchdog_threshold

Dataset Registry

• soup data register — register local datasets by name
• soup data unregister / soup data registry — manage and list registered datasets
• Reference datasets by name in soup.yaml instead of file paths

Security

• HF download: trust_remote_code=False + warning panel
• Download path sanitization and samples cap (1M)
• Registry: name validation, path traversal protection, Rich markup escaping
• Config bounds: watchdog threshold ≤100, patience ≤1000, freeze_layers ≤1000

Stats

• 82 new tests (2061 total, 74 test files)
• 26 files changed, 1936 insertions

Install / Upgrade

pip install --upgrade soup-cli