12 5 months ago

Qwen3.5-4B fine-tuned to generate dbt Silver layer plans (dim_* and fct_* models) from a classified PostgreSQL schema.

ollama run NiChr0/schemalytics-silver-agent

Details

5 months ago

4ef26ecb630b · 2.7GB ·

qwen35
·
4.21B
·
Q4_K_M
{ "num_ctx": 12288, "num_predict": 4096, "temperature": 0 }

Readme

Schemalytics Silver Agent (Agent 4a)

Fine-tuned Qwen3.5-4B for dbt Silver layer plan generation.

Part of the Schemalytics pipeline — a tool that turns a PostgreSQL schema into a full dbt project using a multi-agent LLM pipeline.

What it does

Given a database schema with table classifications (fact/dimension/bridge/reference), this model generates a Silver layer modeling plan:

  • dim_* models for dimension tables (descriptive entities)
  • fct_* models for fact tables (business events/transactions)
  • Selects appropriate grain, measures, and foreign keys for each model

Usage

Used automatically by schemalytics when SCHEMALYTICS_LLM_PROVIDER=ollama:

”`bash ollama pull nichr0/schemalytics-silver-agent SCHEMALYTICS_LLM_PROVIDER=ollama schemalytics generate -c postgresql://… Training Base model: Qwen3.5-4B Method: QLoRA (r=8, α=16) via Unsloth Dataset: ~500 real-world database schemas (Spider benchmark + enterprise schemas) Task: Given schema + table classifications → output structured Silver modeling plan (JSON) Parameters Temperature: 0 (deterministic) Context: 12,288 tokens Max output: 4,096 tokens