ReasonForge
A reinforcement-learning framework for language models that builds adaptive training environments where a model practices problem-solving and is scored with multi-dimensional, partial-credit rewards instead of a single binary right-or-wrong signal.
ReasonForge structures training as a three-phase loop. An environment synthesis phase constructs challenge environments that adapt to the model's current capability level rather than drawing from a fixed static dataset. A reward-shaping phase scores a model's attempt with partial credit, penalty gradients and multi-objective scoring, rather than binary feedback as in standard reinforcement learning from human feedback, so training reflects why an answer's reasoning path was right or wrong and not only whether the final answer matched.
A policy-refinement phase uses proximal policy optimization together with a self-reflection mechanism, where the model critiques its own reasoning and each training iteration records a log of how its reasoning changed.
It is licensed under MIT and requires Python 3.10 or later with a CUDA-compatible GPU.