This is a meeTARA-branded export of the official open-weights checkpoint unknown (Gemma 4, instruction-tuned). Gemma-native turn formatting is preserved, and meeTARA default system behavior is injected only when the caller does not provide a system message, so multimodal formatting, tool calling, and thinking modes stay compatible with Transformers and current llama.cpp converters.

me²TARA = Mental & Emotional Empathetic Technology for Adaptive Response & Assistance.

This folder includes meetara_default_system_en.txt: the same welcome + structured-response intent as the v3 "simple" ChatML block, in plain text for you to pass as a system (or developer) message in your app or API. The same default content is also mirrored into Gemma-native template metadata so behavior remains present when no explicit system message is provided.

convert_to_gguf_meetara_v3_enhanced.py injects a large Qwen-style ChatML chat template. Gemma 4 expects Google's <|turn|> template and special tokens; mixing formats would produce broken prompts. For Gemma, use this script for GGUF packaging; use meetara_default_system_en.txt plus your app, or a Gemma-native fine-tune, if you need meeTARA behavior inside the stack.

Built with repo llama.cpp:

Apache 2.0 (same as upstream Gemma 4).

Weights and tokenizer: Google DeepMind (unknown). Packaging: meeTARA project.