logicless/qwen25-coder-3b-sql-create-context-lora-mlx

Qwen2.5-Coder-3B Text-to-SQL LoRA Adapter for MLX

This is a LoRA adapter for mlx-community/Qwen2.5-Coder-3B-Instruct-4bit, trained locally with MLX-LM on Apple Silicon for schema-conditioned text-to-SQL generation.

Results

On a frozen 1,000-example held-out set with context groups disjoint from the 5,000 training rows:

Metric	Base model	This adapter
Normalized reference-SQL exact match	5.2%	77.0%
Parser-valid SQL	99.0%	99.6%
SELECT-only SQL	99.0%	99.6%

Exact match is normalized agreement with the dataset reference SQL, not a semantic-equivalence or database-execution result. SQL validity is parser-only; no database is opened or executed.

Base model and compatibility

• Base: mlx-community/Qwen2.5-Coder-3B-Instruct-4bit
• Pinned base revision: 3dd939c621c08e5753d5b89f35a2642cd83b98ca
• Runtime tested: MLX 0.32.2 and MLX-LM 0.31.3
• Adapter type: MLX-LM LoRA, rank 16, final 16 transformer layers

This repository contains an adapter only. It requires the base model above and an MLX-LM-compatible Apple Silicon environment.

Usage

from mlx_lm import generate, load

base_model = "mlx-community/Qwen2.5-Coder-3B-Instruct-4bit"
# Replace this placeholder with the local directory containing the downloaded
# files from this Hugging Face repository.
adapter_path = "/path/to/qwen25-coder-3b-sql-create-context-lora-mlx"

model, tokenizer = load(base_model, adapter_path=adapter_path)
messages = [
    {
        "role": "system",
        "content": "You translate natural-language questions into SQLite SQL. Return exactly one read-only SELECT statement. Use only the supplied schema. Return SQL only: no Markdown fences, explanation, or comments.",
    },
    {
        "role": "user",
        "content": "Schema:\\nCREATE TABLE employees (id INTEGER, name TEXT);\\n\\nQuestion: List employee names.\\n\\nSQL:",
    },
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=256, verbose=False))

Training data and method

• Dataset: b-mc2/sql-create-context, revision 9d80a6a118b838d9defc3798d659a54a2ac2ff37
• Inputs: supplied CREATE TABLE schema context plus natural-language question
• Target: reference SQLite SQL answer
• Split: 5,000 training and 1,000 held-out rows grouped by identical full schema context; zero shared context groups
• Training: 2,500 micro-batches, batch size 2, gradient accumulation 4, learning rate 1e-5, prompt loss masking

The complete, runnable training recipe—including project-relative data and adapter paths—is configs/lora.yaml.

Limitations

• This is a local, schema-conditioned subset experiment, not a production SQL agent.
• Exact match can mark semantically equivalent SQL as incorrect.
• The evaluator does not execute queries, validate schema references, or assess result equivalence.
• Use generated SQL with normal application-level authorization and review controls.

License and attribution

The training dataset is distributed under CC-BY-4.0 and should be attributed to b-mc2/sql-create-context. This adapter is derived from Qwen2.5-Coder-3B-Instruct; use is subject to the Qwen Research License. This model card does not grant rights beyond those upstream terms.

Project source

The accompanying experiment code and technical report are published at karyboy/mlx-text-to-sql-lora.