Overview

KeeganC/gemma-3-1b-it-amr_thinking-2 is a 1 billion parameter model based on the Gemma architecture, specifically fine-tuned for generating structured reasoning. It leverages Group Relative Policy Optimization (GRPO) to enhance its ability to produce explicit thought processes. This model is a GRPO-trained version of the chimbiwide/gemma-3-1b-it-thinking-32k-sft-base model, developed using the Tunix (JAX) framework on v6e-1 TPU hardware.

Key Capabilities

• Structured Reasoning Output: Generates a distinct reasoning trace (<reasoning>step-by-step thinking process</reasoning>) before providing the final answer (<answer>final answer</answer>).
• GRPO Training: Utilizes Group Relative Policy Optimization for improved policy learning in reasoning tasks.
• Extended Context Window: Supports a context length of 32768 tokens, allowing for processing longer inputs and more complex reasoning chains.

Good For

• Applications requiring transparent, step-by-step explanations of a model's thought process.
• Tasks where understanding how an answer was derived is as important as the answer itself.
• Developing agents or systems that need to articulate their reasoning for debugging, auditing, or user comprehension.