RajGMore/tinyforge-qwen2.5-3b-titanic-sql-lora

tinyforge: Qwen2.5-3B-Instruct LoRA for text-to-SQL on the Titanic table

A small LoRA adapter (r=16, q_proj/v_proj, 7 MB) trained with tinyforge on a 4 GB RTX 3050 Ti laptop GPU: 100 steps, about 2 minutes, NF4 base with layer streaming. Training data: 1,126 question/SQL pairs generated from the public titanic.csv (891 rows) and verified by executing the SQL.

It answers questions about a table named titanic with a single SQLite query.

Results (read the caveats)

60 held-out questions, greedy decoding, one run. Execution accuracy compares result sets on the real table.

Model	Execution accuracy	Exact match
This adapter	98.3%	98.3%
Base told to "reply with exactly one SQLite query"	70.0%	58.3%
Base with no format instruction	0% (answers in prose)	0%

Caveats: the questions are templated and come from the same table and templates as the training data, so this shows the adapter learned a narrow task, not general SQL skill. A harder-shape test on a different table with the 8B model gave 98% vs 97% (no real gain over a well-prompted base). Raw numbers: eval_results.json.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct", device_map="auto")
model = PeftModel.from_pretrained(base, "RajGMore/tinyforge-qwen2.5-3b-titanic-sql-lora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")

Reproduce the whole pipeline: pip install tinyforge, then see the repo README ("Verified results").