iola-gemma3-router-gemma3-1b-lora

IOLA Gemma 3 Router LoRA

LoRA-адаптер для google/gemma-3-1b-it, обучаемый на датасете LMSerg/iola-gemma3-router-sft.

Цель модели - возвращать строгий JSON/tool-call для CLI и MCP/RAG-слоя Йошкар-Олы, а не отвечать по изменяемым городским данным из памяти.

Expected output

{
  "action": "tool_call",
  "tool": "get_entity_field",
  "args": {
    "layer": "schools",
    "inn": "1215067590",
    "field": "phone"
  }
}

Training

Training script:

training/train_gemma3_router_lora.py

Source repository:

https://github.com/yasg1988/iola-small-model-lab

Dataset:

https://huggingface.co/datasets/LMSerg/iola-gemma3-router-sft

Evaluation

After the first training job completes, this repo should contain:

• adapter files;
• tokenizer files;
• eval/eval_metrics.json;
• eval/eval_predictions.jsonl.

Acceptance target for the first adapter:

• JSON parse: 30/30;
• strict exact match on router-eval-v1: substantially above raw gemma3:1b baseline 2/30.