tdelard/Qwen3-4B-DBT-Instruct-GGUF

Qwen3-4B-DBT-Instruct — GGUF

A fine-tuned version of unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit (the pre-quantized 4-bit Unsloth variant of Qwen3-4B-Instruct) specialized in converting natural language business questions into complete, multi-file dbt DAGs.

Looking for the LoRA adapter? → tdelard/Qwen3-4B-DBT-Instruct-LoRA

---

What does this model do?

Given a business question and a SQL schema, the model generates a full dbt project structure:

• Staging layer — stg_*.sql files that clean and rename raw source data
• YAML sources — _sources.yml / _stg_*.yml schema files with column definitions
• Intermediate models — int_*.sql files that join and enrich staging data
• Marts layer — fct_*.sql or dim_*.sql final business models

Example prompt:

Business question: Show the total revenue per product category, filtered to orders placed in the last 12 months.
SQL context: CREATE TABLE orders (...); CREATE TABLE products (...);

Example output: A ready-to-use dbt DAG with staging, intermediate, and mart SQL + YAML files.

---

Available GGUF files

File	Quantization	Size	Recommended for
qwen3-4b-instruct-2507.Q4_K_M.gguf	Q4_K_M	~2.5 GB	Most users (best size/quality trade-off)
qwen3-4b-instruct-2507.Q5_K_M.gguf	Q5_K_M	~2.9 GB	Higher quality, still fits in 8 GB RAM
qwen3-4b-instruct-2507.Q8_0.gguf	Q8_0	~4.3 GB	Maximum quality, requires ~6 GB RAM

---

Usage

LM Studio / Jan

1. Search for tdelard/Qwen3-4b-DBT-Instruct-GGUF in the model browser, or download the GGUF manually.
2. Load the model and use the system prompt below.

Ollama

An Modelfile is included in this repository for easy import:

ollama create qwen3-dbt -f Modelfile
ollama run qwen3-dbt

llama.cpp

llama-cli -hf tdelard/Qwen3-4b-DBT-Instruct-GGUF --jinja \
  -m qwen3-4b-instruct-2507.Q4_K_M.gguf

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="tdelard/Qwen3-4b-DBT-Instruct-GGUF",
    filename="qwen3-4b-instruct-2507.Q4_K_M.gguf",
    n_ctx=2048,
)

response = llm.create_chat_completion(
    messages=[
        {
            "role": "system",
            "content": (
                "You are a dbt expert. Given a business question and a SQL schema, "
                "generate a complete, production-ready dbt DAG including staging SQL files, "
                "YAML schema files, intermediate models, and mart models. "
                "Use proper dbt conventions: ref(), source(), naming prefixes (stg_, int_, fct_, dim_)."
            ),
        },
        {
            "role": "user",
            "content": (
                "Business question: Show the total revenue per product category.\n"
                "SQL context: CREATE TABLE orders (order_id INT, product_id INT, amount DECIMAL); "
                "CREATE TABLE products (product_id INT, category VARCHAR, name VARCHAR);"
            ),
        },
    ]
)
print(response["choices"][0]["message"]["content"])

---

Training details

Parameter	Value
Base model	unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit
Training framework	Unsloth + TRL SFTTrainer
Method	QLoRA (4-bit quantized base + LoRA adapters)
LoRA rank	32
LoRA alpha	32
Learning rate	2e-4
Max sequence length	2048 tokens
Hardware	Google Colab T4 GPU (15 GB VRAM)
Training dataset	tdelard/text_to_dbt
Train split	900 examples
Eval split	100 examples

Training dataset pipeline

The training data was built from scratch using a synthetic generation pipeline:

1. Source: ~1 000 SQL queries sampled from b-mc2/sql-create-context, filtered and scored on 24 structural complexity features (table count, join depth, aggregation, subqueries…).
2. Generation: Each SQL query was transformed into a multi-file dbt DAG by Claude Sonnet via structured prompting.
3. Validation: Every generated DAG was validated with dbt parse (no database required), catching ref/source resolution errors and YAML issues. Only passing DAGs were kept.

---

Limitations

• Context window is 2 048 tokens — very large schemas or highly complex queries may be truncated.
• The model was trained on single-question → single-DAG examples; multi-model or incremental dbt patterns are not covered.
• Output quality degrades on schemas with many tables (> 8–10); use the intermediate layer to break complexity.

---

License

Apache 2.0 — same as the base Qwen3-4B model.

---

Fine-tuned and converted to GGUF using Unsloth.