gemma3-v12-merged

A merged/fine-tuned checkpoint based on the Gemma 3 architecture (Gemma3ForConditionalGeneration, vision-language, SigLIP vision tower), processed with Unsloth.

This is a private, internal checkpoint used for deploying and testing tool-calling behavior via vLLM on Baseten. It is not intended for general/public use.

Contents

• model.safetensors — merged weights (bf16)
• config.json / generation_config.json — model config
• tokenizer.json / tokenizer_config.json — tokenizer
• preprocessor_config.json — image preprocessor config
• chat_template.jinja — original Gemma 3 chat template (no tool-calling support)
• chat_template_tools.jinja — modified chat template that injects an OpenAI-style tools list into the prompt when present, for tool-calling experiments

Notes

• Native chat template has no tool/function-calling scaffolding (no {% if tools %} block, no <tool_call>-style special tokens in the tokenizer).
• Deployed via a custom Truss model.py wrapping vLLM's AsyncLLMEngine on Baseten (H100).