LMSerg/iola-gemma3-router-sft

IOLA Gemma 3 Router SFT

SFT-набор для первого LoRA-прогона маленькой router-модели на базе gemma3:1b.

Цель датасета - научить модель не отвечать по изменяемым городским данным из памяти, а возвращать строгий JSON/tool-call к источнику истины: FastAPI/MCP, RAG/Qdrant или слой персоналий.

Files

File	Rows	Purpose
data/router-train-v1.jsonl	810	Combined SFT train pack.
data/router-train-entities.jsonl	782	Schools/kindergartens entity routing and adversarial false-premise examples.
data/router-train-safety.jsonl	8	Clarify, refuse, entity search, officials, RAG examples.
data/router-train-history.jsonl	20	Stable city history direct-answer and source-required examples.
data/router-eval-v1.jsonl	30	Held-out eval set, not included in train.
data/router-train-v2.jsonl	810	V2 train pack: model returns entity_number/entity_name; API resolves INN.
data/router-eval-v2.jsonl	30	V2 held-out eval set.
data/router-train-v3.jsonl	829	V3 train pack with extra alias/search/clarify/refuse/RAG/officials examples.
data/router-eval-v3.jsonl	30	V3 held-out eval; alias queries expect entity_name.
data/router-train-v4.jsonl	856	V4 train pack with targeted fixes for V3 failures.
data/router-eval-v4.jsonl	30	V4 held-out eval on the V3 contract surface.
data/router-train-v5.jsonl	890	V5 train pack with targeted fixes for V4 failures.
data/router-eval-v5.jsonl	30	V5 held-out eval on the V4 contract surface.
data/router-train-v6.jsonl	910	V6 train pack with salary/private-data refusal reinforcement.
data/router-eval-v6.jsonl	30	V6 held-out eval on the V5 contract surface.

Format

Each row is JSONL with two chat messages:

{
  "id": "router-train-entity-simple-002-01",
  "dataset": "router-train-entities",
  "source_id": "simple-002",
  "type": "entity_field",
  "messages": [
    {
      "role": "user",
      "content": "дай телефон школы № 2"
    },
    {
      "role": "assistant",
      "content": "{\"action\":\"tool_call\",\"tool\":\"get_entity_field\",\"args\":{\"layer\":\"schools\",\"inn\":\"1215067590\",\"field\":\"phone\"}}"
    }
  ],
  "tags": ["entity_field", "schools", "phone"]
}

For schools and kindergartens, assistant messages do not contain final facts such as phone numbers, addresses, websites, or head names. They contain only routing decisions.

V2 avoids forcing the 1B model to memorize arbitrary INNs:

{
  "action": "tool_call",
  "tool": "resolve_entity_field",
  "args": {
    "layer": "schools",
    "entity_number": 2,
    "field": "phone"
  }
}

Contract

Allowed actions:

• tool_call
• clarify
• refuse
• direct_answer

Allowed tools:

• get_entity_field
• search_entities
• rag_search
• get_current_official
• get_official_by_date

The full contract and validators are maintained in the GitHub repository:

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

Validation

The source repository validates:

• strict assistant JSON;
• allowed actions, tools, layers and fields;
• get_entity_field arguments;
• no exact train/eval question overlap.

Commands:

npm run generate:router
npm run validate:router
npm test