voidful/gemma-4-12b-it-agent-sft-tw-fullft
The voidful/gemma-4-12b-it-agent-sft-tw-fullft model is a 12 billion parameter, full-weight fine-tune of Google's Gemma-4-12B-it. This model is specifically optimized for agentic tasks, having been fine-tuned on the voidful/agent-sft dataset. It demonstrates improved performance over its base model on the CLAW TW core evaluation, making it suitable for agent-based applications requiring enhanced task execution.
Loading preview...
Model Overview
This model, voidful/gemma-4-12b-it-agent-sft-tw-fullft, is a 12 billion parameter language model derived from google/gemma-4-12B-it. Unlike LoRA or QLoRA adaptations, this is a full-weight fine-tune, meaning the language model weights were directly updated during training, while vision/audio embedding stacks remained frozen to preserve multimodal capabilities.
Key Characteristics
- Full Fine-Tune: Direct update of language model weights, not an adapter or merged LoRA.
- Agentic Optimization: Fine-tuned on the
voidful/agent-sftdataset, enhancing its performance for agent-based tasks. - Multimodal Interface Preserved: The original multimodal capabilities of the base Gemma-4 model are maintained.
- Improved Performance: Achieves a CLAW TW core score of 9.567 / 20 (47.83%), outperforming the baseline
google/gemma-4-12B-itwhich scored 8.106 / 20 (40.53%) under the same evaluation setup. - Context Length: Utilizes a sequence length of 8192 during training.
Use Cases
This model is particularly well-suited for:
- Agent-based applications: Where enhanced task execution and understanding are critical.
- Applications requiring improved performance over base Gemma-4-12B-it: Especially in scenarios similar to the CLAW TW core evaluation.
- Multimodal tasks: Where the preserved vision/audio capabilities of the base model are beneficial alongside improved language understanding.