Hey makers! I just launched LoomMint, an open-source, local-first toolkit designed to take the guesswork and security risks out of Supervised Fine-Tuning (SFT) workflows.

If you have ever accidentally executed bad data scripts or struggled to align raw JSON/CSV files into proper conversational formats, LoomMint is built for you. It inspects, validates, and canonicalizes your datasets locally, pairs them with a bounded AI trainer/planner, and runs SFT via TRL or Unsloth—all without generating or executing arbitrary code.

🚀 Get Started Instantly:

bash

pip install loommint
# For training support: pip install loommint[trl] or loommint[unsloth]

Use code with caution.

Key Features:

Safe Data Inspection: Maps ShareGPT, Alpaca, Prompt/Completion, and Q&A layouts without displaying or leaking raw text row contents.
Strict Validation: Zero silent failures. Malformed or assistant-less records are rejected outright instead of guessing.
Bounded AI Planner: Safely generates training blueprints and dataset mappings from structural metadata alone.
Local Execution: Runs supervised fine-tuning (TRL SFT + LoRA / Unsloth QLoRA) completely under your manual approval.

Check out the repository, drop a star, and let me know your thoughts!

📦 PyPI: https://pypi.org/project/loommint/
🛠️ GitHub: https://github.com/hskapasi/loomMint