README.md

---
license: gemma
base_model: google/functiongemma-270m-it
tags:
  - gguf
  - function-calling
  - tool-use
  - llama.cpp
---

# Athena FunctionGemma-270M (fine-tuned)

LoRA fine-tune of `google/functiongemma-270m-it` for Athena, the on-device HR
assistant in [Mis-Genie](https://github.com/MAGSoft/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.