trjxter/Gwimi-4-12B-IT-BF16
trjxter/Gwimi-4-12B-IT-BF16 is a 12 billion parameter Gemma 4 instruction-tuned model, optimized for reasoning tasks. It underwent Supervised Fine-Tuning (SFT) on a 20,000-example reasoning dataset and subsequent Group Sequence Policy Optimization (GSPO) using programmatic correctness, formatting, and anomaly rewards. This BF16 model excels at mathematical, scientific, and technical reasoning, coding, and debugging, with a context length of 32,768 tokens.
Loading preview...
Gwimi-4-12B-IT-BF16: A Reasoning-Optimized Gemma 4 Model
Gwimi-4-12B-IT-BF16 is a 12 billion parameter instruction-tuned model based on the Gemma 4 architecture, developed by trjxter. It has been rigorously optimized through a two-stage post-training process to enhance its reasoning capabilities.
Key Capabilities & Training
- Supervised Fine-Tuning (SFT): The model was initially fine-tuned on a curated dataset of 20,000 reasoning examples, covering mathematics, science, coding, debugging, and general technical reasoning. This stage used a rank-128 rsLoRA adapter with a maximum sequence length of 32,768 tokens.
- Group Sequence Policy Optimization (GSPO): Following SFT, the model underwent GSPO on 12,000 reinforcement learning prompts. This phase utilized programmatic reward functions for correctness, formatting, and anomaly detection, specifically targeting verifiable mathematics, science, and structured reasoning.
- Merged BF16 Weights: The final release integrates the cumulative SFT and GSPO LoRA updates directly into the original BF16 base weights, eliminating the need for a separate adapter during inference.
Intended Uses
This model is designed for experimentation and research in areas such as:
- Mathematical and scientific reasoning
- Technical question answering
- Coding and debugging assistance
- Long-form structured reasoning and instruction following
- Local inference research and comparison of SFT-only vs. SFT+RL pipelines.
It is important to note that while SFT evaluation metrics are provided, no independent final benchmark suite is reported yet, and the model's performance may vary outside its primary optimization distributions.