Model Overview

This model, developed by Phonsiri Thabunsri and CYP777, is a supervised fine-tuned (SFT) version of Google's gemma-2-2b-it base model. Its primary purpose is to generate structured chain-of-thought reasoning for mathematical and logical problems, explicitly separating the reasoning process from the final answer.

Key Capabilities

• Structured Reasoning: Produces detailed, step-by-step reasoning enclosed within <reasoning>...</reasoning> tags, followed by the concise final answer in <answer>...</answer> tags.
• Foundation for Downstream Models: Serves as the base for further training, such as the Phonsiri/gemma-2-2b-GRPO-Reasoning-full model.
• Full Parameter Fine-tuning: Unlike many models using LoRA/PEFT, this model underwent full parameter fine-tuning for comprehensive adaptation.

Training Details

The model was trained for 3 epochs with a learning rate of 2e-5 and a maximum sequence length of 8192. Training data included the nohurry/Opus-4.6-Reasoning-3000x-filtered dataset, alongside several local datasets containing Thai and general mathematical problems and solutions.

Good for

• Applications requiring transparent and verifiable problem-solving steps.
• Educational tools that need to show how a solution is derived.
• Developing agents that can explain their decision-making process in a structured format.
• As a base model for further reinforcement learning or fine-tuning on reasoning tasks.