Qwen3.5-0.8B Text-to-SQL LoRA adapter (MLX)

A LoRA adapter for mlx-community/Qwen3.5-0.8B-bf16 that turns a natural-language question plus a CREATE TABLE schema into a single SQLite query. Trained and run with Apple MLX (mlx-lm).

Training status: completed.

Model overview

Model details

• Base model: mlx-community/Qwen3.5-0.8B-bf16 at commit 3067585164dbcc505eec73d554349de6a27571a4, an MLX bf16 conversion of Qwen/Qwen3.5-0.8B. Its license is apache-2.0. The base repository is a vision-language checkpoint; MLX-LM loads only its text model, which is what was fine-tuned.
• This repository contains only a LoRA adapter (adapters.safetensors + adapter_config.json). It is not a standalone model: it must be loaded on top of the exact base model above.
• Weight format: safetensors (MLX). Base weights are bf16, not quantized; the adapter was trained against the bf16 base.
• Parameters: trainable 3.608M of 752.392M total (as reported by MLX-LM during training).
• Software: mlx 0.32.4, mlx-lm 0.32.2, huggingface-hub 2.3.0.

Intended use

Converting an English question about one or more tables, together with their CREATE TABLE statements, into a SQLite query. Suitable for experiments, teaching and studying small-model fine-tuning on a laptop.

Generated SQL must be reviewed or validated before it is executed. Do not run model output against production or sensitive databases without read-only credentials, query allow-listing and resource limits.

Out of scope: other SQL dialects, schemas without CREATE TABLE context, data-modifying statements, and any use where incorrect SQL causes harm.

Training data

• Dataset: b-mc2/sql-create-context, file sql_create_context_v4.json, revision 9d80a6a118b838d9defc3798d659a54a2ac2ff37, license cc-by-4.0.
• Content: 78,577 English questions, each with CREATE TABLE context and a reference SQL answer; built by its author from WikiSQL and Spider.
• Source rows: 78,577. The source has no predefined splits.
• Split: one shuffle of all rows with random.Random(42), then test = first 100, validation = next 200, train = next 2000. The remaining rows are not used.
• Prepared sizes actually used: test 100, valid 200, train 2000.
• Filtering: only malformed rows (missing/empty fields) would be dropped; dropped: test 0, valid 0, train 0.
• Deduplication: exact (question, schema) overlap between splits is measured and reported: test/valid 0, test/train 0, valid/train 0. Dropping is disabled.
• Prompt (user turn):

{schema}

Question: {question}
Write a SQLite query that answers the question. Return only the SQL query.

• Target (assistant turn): the dataset's SQL answer, verbatim.
• Format: MLX-LM chat JSONL ({"messages": [user, assistant]}); loss is computed on the assistant answer only (mask_prompt).
• Limitations of the data: schemas carry column types inferred by the dataset author and no rows; many WikiSQL-derived questions are single-table lookups; values are often lower-cased; some references are themselves imperfect.

Training procedure

Evaluation

Held-out test split (100 examples, never used for training or validation). Greedy decoding, thinking disabled, at most 200 new tokens. The SQL is taken from the first ```sql block or the first SELECT in the answer.

Metrics

• Execution accuracy (primary): the schema is instantiated in 3 in-memory SQLite databases with 25 seeded random rows per table (biased towards the literals in the reference query). A prediction is correct when it returns the same multiset of rows as the reference on every database. Model SQL runs read-only, with a time limit, in a sandbox without file or network access.
• Exact match: identical text. Normalized exact match: identical after lower-casing outside string literals, unifying quotes and whitespace.
• Valid SQL rate: the query runs read-only against the empty schema. Valid SQL is not necessarily correct SQL.

Random-row execution checks can accept a wrong query that happens to return the same rows on the synthetic data. 100 examples give wide uncertainty: a 95% interval is up to about ±10 percentage points.

How to use

Requires macOS on Apple Silicon.

pip install "mlx-lm>=0.32.2,<0.33" huggingface_hub

from huggingface_hub import snapshot_download
from mlx_lm import generate, load

adapter_dir = snapshot_download("thealper2/mlx-qwen3.5-0.8b-text-to-sql")
model, tokenizer = load(
    "mlx-community/Qwen3.5-0.8B-bf16",
    revision="3067585164dbcc505eec73d554349de6a27571a4",
    adapter_path=adapter_dir,
)

schema = "CREATE TABLE head (age INTEGER)"
question = "How many heads of the departments are older than 56 ?"
prompt = (
    f"{schema}\n\nQuestion: {question}\n"
    "Write a SQLite query that answers the question. Return only the SQL query."
)
text = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    tokenize=False,
    enable_thinking=False,
)
print(generate(model, tokenizer, prompt=text, max_tokens=200))

Use exactly this prompt layout and enable_thinking=False; the adapter was trained on it. mlx_lm.load applies the adapter from a local folder, which is why the repository is downloaded first.

Limitations and risks

• Generates incorrect SQL regularly, including queries that run but answer a different question.
• May misread the schema or hallucinate table and column names, especially for schemas unlike the training distribution (WikiSQL/Spider-style tables).
• Targets SQLite; other dialects (PostgreSQL, MySQL, T-SQL) may need changes.
• Ambiguous questions get one arbitrary interpretation; there is no clarification step.
• Trained on 2,000 examples: behaviour on long multi-table schemas, nested queries and unseen domains is weak and not well measured.
• Evaluation uses synthetic rows and a small test set; see the Evaluation section.
• Small model on consumer hardware: no production-readiness, latency or robustness guarantees.
• Never execute generated SQL against sensitive databases without validation, least-privilege read-only access and resource limits.

License and attribution

• Adapter: released under apache-2.0 (as configured by the publisher).
• Base model: mlx-community/Qwen3.5-0.8B-bf16 / Qwen/Qwen3.5-0.8B, license apache-2.0.
• Dataset: b-mc2/sql-create-context, license cc-by-4.0, which requires attribution. It was built from WikiSQL (Zhong et al., 2017, Seq2SQL) and Spider (Yu et al., 2018); see the dataset card for its citation.
• The code, dataset and model licenses are independent; check each before reuse.

Reproducibility

The project that produced this adapter keeps every setting in configs/ (data.yaml, training.yaml, evaluation.yaml, publishing.yaml) and drives everything through make:

make install
make data-prepare        # download the pinned dataset file, build the splits
make data-validate
make train PROFILE=m1_air_8gb
make evaluate-base       # baseline
make evaluate            # base + adapter
make card
make publish-dry-run

Artifacts: adapter in artifacts/adapters/m1_air_8gb, run manifest run_manifest.json (configuration, software versions, losses, timings), evaluation records under reports/evaluation/. Hardware: an Apple Silicon Mac; the m1_air_8gb profile targets an 8 GB M1 MacBook Air.