1010happy/AmongUsModels

Overview

1010happy/AmongUsModels is a 1.5 billion parameter language model, fine-tuned from the Qwen/Qwen2.5-1.5B base model. It leverages a substantial 32768-token context window, allowing for processing longer inputs and maintaining coherence over extended interactions. The model's training utilized the TRL framework, specifically incorporating the GRPO (Gradient-based Reasoning Policy Optimization) method.

Key Capabilities

• Enhanced Mathematical Reasoning: The core differentiator of this model is its training with GRPO, a method introduced in the research paper "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models" (arXiv:2402.03300). This indicates a specialization in handling complex mathematical problems and logical reasoning tasks.
• Qwen2.5 Foundation: Built upon the Qwen2.5 architecture, it inherits the general language understanding and generation capabilities of its base model.
• Long Context Handling: With a 32768-token context length, it can process and generate responses based on extensive input, beneficial for detailed problem descriptions or multi-turn conversations.

Training Details

The model was fine-tuned using TRL (Transformers Reinforcement Learning) and the GRPO method. This approach suggests a focus on improving specific performance metrics, particularly in reasoning, through reinforcement learning techniques.

Good For

• Applications requiring strong mathematical problem-solving.
• Tasks that benefit from advanced logical reasoning.
• Scenarios where processing long and detailed prompts is crucial.