GitHub Repository

Fine-tune Qwen3.5 for Text-to-SQL on your Mac with MLX LoRA — no GPU required. 4-model comparison, prompt engineering benchmark, and honest limitations.

6 starsPython

I fine-tuned Qwen 3.5 (0.8B–4B) on a Mac for text-to-SQL – 2B beats 12B

Name: I fine-tuned Qwen 3.5 (0.8B–4B) on a Mac for text-to-SQL – 2B beats 12B
Availability: InStock
Author: sciences44

by sciences44·Mar 5, 2026·7 points·1 comment

Visit Project View on HN

AI Analysis

●●SolidShip ItNiche GemBig Brain

Unified memory trick lets a 2B model beat 12B; trains on MacBook with zero cloud costs.

Strengths

•Honest limitations section and methodical 4-model comparison with identical hyperparameters.
•Minimal setup: uv sync → train → eval removes friction from typical cloud-based fine-tuning.
•Exploits Apple Silicon's unified memory architecture—a non-obvious technical insight most devs miss.

Weaknesses

•Text-to-SQL fine-tuning is well-trodden territory; no novel task or dataset contribution.
•50% semantic accuracy on evaluation task leaves the model's real-world utility unclear.