codelion/pts

Identifies decision-point tokens in LLM generations that significantly impact task success, then converts them into DPO fine-tuning datasets or steering vectors for inference optimization. Supports multiple reasoning formats (chain-of-thought, GSM8k, MATH) and includes a web visualizer for exploring token-level reasoning circuits and causal dependencies. Integrates with Hugging Face Hub for dataset distribution and works with activation-based steering approaches like optillm.

146 stars.