WindyITS/business-model-kg-query-stack

Business Model KG Query Stack

This repository contains the deployable local query stack for Business Model KG. It is not a standalone chatbot model. It is a runtime bundle used by the project to decide how a question should be handled and, when safe, to produce a structured local query plan.

The bundle combines:

• a fine-tuned DeBERTa-v3-small router classifier
• calibrated router thresholds
• an MLX QLoRA planner adapter for mlx-community/Qwen3-4B-Instruct-2507-4bit
• the frozen planner system prompt
• the runtime bundle manifest

The package intentionally excludes optimizer states and intermediate training checkpoints. It is the deployable bundle, not the full training workspace.

Layout

manifest.json
router/
  thresholds.json
  model/
planner/
  system_prompt.txt
  adapter/

Components

Router:

microsoft/deberta-v3-small

The router is stored as a full fine-tuned classifier under router/model/. It predicts whether a question is suitable for local answering, should fall back to a stronger hosted model, or should be refused. Runtime thresholds are stored in router/thresholds.json.

Planner:

mlx-community/Qwen3-4B-Instruct-2507-4bit

The planner is stored as an adapter only under planner/adapter/. To use it, the runtime also needs the base model above. The planner system prompt is included in planner/system_prompt.txt.

Runtime Role

When the router selects the local path, the planner emits a compact supported query plan. The Business Model KG Python runtime then validates that plan and compiles it into read-only Cypher. This separation keeps the planner constrained to the project ontology and avoids letting it execute arbitrary database logic.

Intended Use

Use this repository as the contents of:

runtime_assets/query_stack/

inside the Business Model KG project. The expected file locations are described by manifest.json.

Training Data

The related training dataset is published separately as:

WindyITS/business-model-kg-query-planner-data

The router was trained on 8,000 examples and validated on 1,200 examples across the labels local, api_fallback, and refuse.

No standalone license has been selected for this release yet. Check the upstream terms for the base models before redistribution or production use.