AmbikaSoni/qwen2.5-coder-1.5b-sql-lora

Qwen2.5-Coder-1.5B Text-to-SQL (QLoRA fine-tune)

LoRA adapter for Qwen2.5-Coder-1.5B-Instruct, fine-tuned on text-to-SQL generation.

Results

Evaluated on 100 held-out examples from b-mc2/sql-create-context (exact-match after SQL normalization):

Model	Baseline	Fine-tuned	Δ
Qwen2.5-1.5B-Instruct (V1)	42.0%	61.0%	+19.0
Qwen2.5-Coder-1.5B-Instruct (this)	42.0%	69.0%	+27.0

The Coder base model doesn't score higher out-of-the-box on this exact-match eval, but fine-tunes to a higher ceiling (+8 points over vanilla Qwen with identical LoRA config and training data).

Training details

• Method: QLoRA (4-bit NF4 quantization + LoRA)
• Base: Qwen/Qwen2.5-Coder-1.5B-Instruct
• Data: 2,500 examples from b-mc2/sql-create-context, 1 epoch
• LoRA: r=16, α=32, dropout=0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Trainable params: ~18M (1.18% of total)
• Optimizer: paged_adamw_8bit, LR 2e-4, cosine schedule, 5 warmup steps
• Effective batch size: 16
• Hardware: Tesla T4 (Google Colab free tier), ~28 min training

Ablation: doubling LoRA rank did not help

A parallel run with r=32, α=64 and everything else identical gave exactly 69.0% test accuracy — no gain. Training loss was slightly worse at every step, suggesting the extra adapter capacity added init noise without adding useful representational power at this data volume. Adapter is available at AmbikaSoni/qwen2.5-coder-1.5b-sql-lora-r32.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "AmbikaSoni/qwen2.5-coder-1.5b-sql-lora")
tokenizer = AutoTokenizer.from_pretrained("AmbikaSoni/qwen2.5-coder-1.5b-sql-lora")

prompt = '''### Instruction:
Given the schema, write a SQL query to answer the question.

### Schema:
CREATE TABLE employees (id INT, name VARCHAR, department VARCHAR, salary INT)

### Question:
What is the average salary in the Engineering department?

### SQL:
'''

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=100, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Limitations

• Only 2,500 training examples — struggles with complex JOINs, multi-table subqueries, and less common SQL patterns
• Learned dataset-specific formatting conventions (double quotes for string literals, lowercase values, unicode dashes). May not match your target SQL dialect exactly.
• English only
• Small model (1.5B) — for production, consider larger bases

Author

Fine-tuned by Ambika as a learning project. Part of studying QLoRA and efficient fine-tuning at IIT Bombay.