marshadbits/athena-functiongemma-270m

Athena FunctionGemma-270M (fine-tuned)

LoRA fine-tune of google/functiongemma-270m-it for Athena, the on-device HR assistant in Mis-Genie. Text query -> structured tool call (name + arguments) only -- never free-form text.

• Real-eval launch gate (hand-written, naturally-phrased HR queries, distinct from training data): 18/20 correct tool+arguments.
• Synthetic held-out eval: 43/46.
• Trained on 458 examples covering 7 read-only data tools + a clarification- popup fallback for underspecified queries.

GGUF file (athena-functiongemma-270m-f16.gguf) is ready to serve directly with llama-server (raw llama.cpp, not Ollama) -- pass an explicit "stop": ["<end_function_call>"] in requests; llama-server has no built-in parser for this model's tool-call syntax and will otherwise keep generating past the call.

Two known, understood gaps (not fixed by more training data):

• Partial employee names ("Taylor" -> "Taylor Reyes") aren't resolved to full names -- an entity-resolution problem for the calling application, not the model.
• One specific day-range phrasing ("between March 1st and March 15th") is under-represented in training (12/458 examples) and occasionally fails to generate a parseable call.