JinyiHan/JET-1.5B
JinyiHan/JET-1.5B is a 1.5 billion parameter language model, based on the DeepSeek-Distill-Qwen-1.5B architecture, designed for efficient reasoning. It utilizes a reinforcement learning framework to generate high-quality reasoning steps while minimizing computational cost and token usage. With a 32768 token context length, this model is optimized for tasks requiring structured thought processes and reduced inference overhead.
Loading preview...
Overview
JinyiHan/JET-1.5B is a 1.5 billion parameter language model developed by JinyiHan. It is built upon the DeepSeek-Distill-Qwen-1.5B base model and has been fine-tuned using a reinforcement learning (RL) framework. The primary goal of this training is to enhance the model's reasoning capabilities while simultaneously optimizing for efficiency by reducing unnecessary computation and token consumption.
Key Capabilities
- Efficient Reasoning: The model is specifically trained to generate high-quality reasoning steps with a focus on minimizing resource usage.
- RL-based Fine-tuning: Leverages a reinforcement learning framework to improve its reasoning process.
- Optimized Token Usage: Designed to reduce the number of tokens required for complex reasoning tasks.
Training Details
The training pipeline, named Just-Enough-Think, is publicly available on GitHub. This repository includes scripts for:
- RL-based fine-tuning procedures.
- Evaluation and benchmarking of the model's performance.
Chat Template
The model uses a specific chat template that includes a system prompt guiding the assistant to solve user questions and wrap the final answer in \boxed{}. This structured prompt ensures consistent output formatting for reasoning tasks.