W-Kaski/qwen25-15b-text2sql-lora-r32

Qwen2.5-1.5B Text-to-SQL QLoRA Adapter Rank 32

This adapter was trained for a reproducible Text-to-SQL experiment using Qwen/Qwen2.5-1.5B-Instruct as the base model.

Intended Use

This adapter is intended for controlled Text-to-SQL experiments and demos where the user provides an explicit SQL schema and one natural-language question.

It is not intended for production analytics, arbitrary database access, safety-critical workflows, or unsandboxed query execution.

Training Setup

• base model: Qwen/Qwen2.5-1.5B-Instruct
• dataset: b-mc2/sql-create-context
• source fields: answer, question, context
• training rows: 5000
• eval rows: 500
• method: QLoRA
• quantization: 4-bit NF4
• LoRA rank: 32
• LoRA alpha: 64
• LoRA dropout: 0.05
• epochs: 1
• max sequence length: 1024

Results

Evaluation on the 500-row Text-to-SQL split:

Metric	Value
Exact Match	0.712
SQL parse valid	0.990

Deployment-style SQLite evaluation:

Metric	Value
cases	30
parse valid rate	1.000
select-only rate	1.000
execution-valid rate	1.000
execution accuracy	0.600

Limitations

The adapter is better at generating SQL-shaped output and matching dataset-specific SQL patterns than the base model, but it is not a general SQL assistant.

Known failure modes:

• case-sensitive value mismatch
• wrong selected column
• wrong string predicate
• imperfect GROUP BY and HAVING
• imperfect NULL handling
• LIMIT/OFFSET mistakes

Use parse validation, read-only checks, execution sandboxing, timeout protection, and row limits before exposing generated SQL to users.

Repository

Project repo: https://github.com/W-Kaski/qwen-qlora-sql-benchmark