neural-txt
A Python library that runs structured natural-language tasks, such as extracting bullet points, generating question-and-answer pairs and building knowledge graphs, locally through a fine-tuned 135-million-parameter language model instead of a large general-purpose one.
neural-txt is built around a small, fine-tuned language model intended for narrow, structured NLP tasks on technical text, running cheaply enough for resource-constrained environments, with either a Hugging Face and PyTorch backend or an Apple Silicon MLX backend. A reasoning variant of the model emits its reasoning inside a delimited block before the final answer; by default the library strips that block for plain-text output, but it can also return the reasoning and raw model text alongside the answer, including across multiple sampled rollouts.
A companion scoring model, NeuralTxtReward, scores a generated response against a reference answer, and can score a single answer, score a batch, or rank several candidate responses against each other.
It installs from PyPI with optional Hugging Face or MLX extras, and is licensed under MIT.