KDDSTLC/lfed-qwen2.5-coder-14b-sql-lora-warehouse-r64

LFED SQL Assistant v2 — Qwen2.5-Coder-14B-LoRA (Warehouse r=64)

Trained on the real DuckDB warehouse schema (main_core, main_analytics, main_staging — ~30+ tables) for the Local First Education Data Framework.

Differences from v1

	v1 (r=32)	v2 (r=64)
Training schema	5 synthetic tables	Real warehouse (~30 tables)
LoRA rank	32	64
Adapter size	~551 MB	~1.1 GB
Target tables	students, enrollment, attendance, discipline, grades	main_core., main_analytics., main_staging.*

Intended use

Converts natural-language questions about K-12 school data into read-only DuckDB SQL queries. Designed for the LFED Gradio app.

Training data

Synthetic NL→SQL pairs generated against the real warehouse schema using Qwen2.5-72B-Instruct-AWQ via vLLM on Modal A100-80GB. Every generated SQL was validated by executing it against the real DuckDB warehouse.

How to use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/qwen2.5-coder-14b-instruct-bnb-4bit"
adapter = "KDDSTLC/lfed-qwen2.5-coder-14b-sql-lora-warehouse-r64"

model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, adapter, torch_device="cpu")
tokenizer = AutoTokenizer.from_pretrained(adapter)

Known limitations

• Trained on synthetic data against the real schema; may hallucinate columns
• Does not know about Rill metrics views (see project docs for the known gap)
• Requires CUDA for 4-bit inference (bnb); CPU/MPS needs a different base model