NohTow/PPL-MCTS

Implements Monte Carlo Tree Search decoding that combines any Hugging Face language model with a discriminator to enforce constraints during generation, using configurable exploration parameters (c_puct, temperature, repetition penalty) and rollout-based value estimation. The approach supports plug-and-play discriminators from custom PyTorch classifiers to vanilla transformers, with optimizations for unidirectional attention models to enable cached hidden state reuse and significant speedup gains.

No commits in the last 6 months.