function-gemma-finetuned-tool-call

Fine-tuned Function-Gemma 270M model for bilingual (English/French) tool-calling.

Files

• function-gemma-finetuned-tool-call.gguf (F16 merged GGUF)

Base Model

• unsloth/functiongemma-270m-it

Training Summary

• Method: SFT + LoRA, then merged into full weights
• Dataset: custom bilingual EN/FR tool-calling set (dataset_80tools_en_fr.json)
• Target behavior: structured function/tool calls with argument extraction and no-tool abstention when appropriate

Local Evaluation (checkpoint benchmark)

From outputs/eval_checkpoint_report.json:

• Total cases: 16
• Pass rate: 0.8125
• Decision accuracy: 0.8125
• Tool name accuracy: 0.8125
• Argument presence accuracy: 1.0
• Tool-call recall: 1.0
• No-tool precision: 0.5

Usage (llama.cpp)

llama.cpp/build/bin/llama-cli \
  --model function-gemma-finetuned-tool-call.gguf \
  --ctx-size 32768 \
  --n-gpu-layers 99 \
  --seed 3407 \
  --top-k 64 \
  --top-p 0.95 \
  --temp 1.0 \
  --jinja

For one-shot test:

llama.cpp/build/bin/llama-cli \
  --model function-gemma-finetuned-tool-call.gguf \
  --ctx-size 32768 \
  --n-gpu-layers 99 \
  --seed 3407 \
  --top-k 64 \
  --top-p 0.95 \
  --temp 1.0 \
  --jinja \
  --single-turn \
  --simple-io \
  --prompt "What is the weather in Paris?"

Prompt / Output Format

This model was fine-tuned for Function-Gemma style tool tags (e.g. <start_function_call>...). When used with --jinja, llama.cpp applies the chat template stored in GGUF metadata.

Limitations

• Small model (270M): can still over-call tools in ambiguous no-tool prompts.
• Best results require strong tool schema prompts and clear user intent.

Intended Use

• Lightweight local assistant prototypes
• Tool-routing and structured argument extraction tasks
• EN/FR bilingual demos and experimentation