Word2Li/CoTranslator-4B

CoTranslator-4B: Reasoning Trajectory Refiner

CoTranslator-4B is a 4-billion parameter model developed by Word2Li, based on the Qwen3-4B-Thinking-2507 architecture. Its core function is to act as a "reasoning trajectory refiner," translating complex, answer-conditioned reverse chain-of-thought (CoT) trajectories into more natural, verified forward reasoning paths. This process is crucial for generating high-quality synthetic supervision data, which can then be used to train smaller, more efficient student models for reasoning tasks.

Key Capabilities

• Reverse to Forward CoT Translation: Converts detailed reverse CoT trajectories (generated with access to an answer) into trajectories resembling standard forward solutions.
• Reasoning Distillation: Facilitates the distillation of complex reasoning abilities from larger teacher models into smaller student models by providing structured training data.
• Mathematical Reasoning Support: Specifically trained on a subset of paired mathematical reasoning data, making it adept at refining mathematical problem-solving steps.
• Supervised Fine-Tuning: Utilizes full-parameter supervised fine-tuning on a dataset of 33,286 paired reverse and forward trajectories.

Good For

• Generating Training Data: Ideal for researchers and developers looking to create synthetic, high-quality reasoning trajectories to supervise the training of other language models.
• Improving Student Model Performance: Useful for distilling advanced reasoning capabilities into smaller, more efficient models without requiring extensive human annotation.
• Mathematical Problem Solving: Particularly effective in contexts requiring the refinement or generation of structured reasoning for mathematical problems.
• Research in Reasoning Distillation: A valuable tool for exploring and advancing techniques in reasoning distillation and chain-of-thought prompting.