Michael-Kozu/rlang-reasoning-traces

RLang Reasoning Traces是一个结构化推理痕迹的数据集，使用RLang（一种受Rust启发的合成推理语言）设计，旨在用类型化、压缩、可验证的推理替代冗长的英语思维链。该数据集训练模型默认使用RLang进行思考，实现比传统英语思维链3-5倍的标记压缩。数据集包含256个示例，涵盖数学、代码、代数、推理等多个领域，分为不同难度级别。每个示例采用消息格式，兼容HF TRL SFTTrainer。数据集还提供了质量指标、RLang语言概述和训练建议。

RLang Reasoning Traces is a dataset of structured reasoning traces in RLang, a Rust-inspired synthetic reasoning language designed to replace verbose English chain-of-thought with typed, compressed, verifiable reasoning. The dataset trains models to think in RLang by default, achieving 3–5× token compression over conventional English CoT. It contains 256 examples covering various domains like math, code, algebra, and reasoning, with different difficulty levels. Each example is in messages format, compatible with HF TRL SFTTrainer. The dataset also includes quality metrics, an overview of RLang, and training recommendations.