GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
PratikshaMahajan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
PratikshaMahajan
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
This is a LoRA adapter for Qwen2.5-1.5B-Instruct that I trained to write SQL from a table schema and a plain-English question. I built it as a learning project to go through the full fine-tuning cycle: preparing data, training, measuring before and after, and publishing.
Code and training notebook: github.com/PratikshaMahajan15/text2sql-finetune
I tested on 200 questions that were not used in training. A prediction counts as correct if it matches the reference SQL after light cleanup (lowercase, extra whitespace removed, double quotes turned into single quotes, markdown fences and trailing semicolons stripped).
| Model | Exact-match accuracy |
|---|---|
| Qwen2.5-1.5B-Instruct (base) | 46.0% |
| Base + this adapter | 68.0% |
Most of the base model's misses come from formatting, such as wrapping the query in a markdown block. The adapter mainly teaches the model to output only the SQL in the style of the dataset.
b-mc2/sql-create-context (seed 42), split into 1,800 for training and 200 for testingpeft and trl's SFTTrainerpython
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigfrom peft import PeftModel base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "PratikshaMahajan/qwen2.5-1.5b-text2sql-qlora" tok = AutoTokenizer.from_pretrained(base_id)bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16)model = AutoModelForCausalLM.from_pretrained(base_id, quantization_config=bnb, device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [ {"role": "system", "content": "You are an expert SQL writer. Given a table schema and a question, reply with only the SQL query."}, {"role": "user", "content": "Schema:\nCREATE TABLE head (age INTEGER)\n\nQuestion: How many heads are older than 56?"},]prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tok(prompt, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))Base model: Qwen2.5-1.5B-Instruct. Dataset: b-mc2/sql-create-context, which is built from WikiSQL and Spider.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfigfrom peft import PeftModel base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "PratikshaMahajan/qwen2.5-1.5b-text2sql-qlora" tok = AutoTokenizer.from_pretrained(base_id)bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16)model = AutoModelForCausalLM.from_pretrained(base_id, quantization_config=bnb, device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [ {"role": "system", "content": "You are an expert SQL writer. Given a table schema and a question, reply with only the SQL query."}, {"role": "user", "content": "Schema:\nCREATE TABLE head (age INTEGER)\n\nQuestion: How many heads are older than 56?"},]prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tok(prompt, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))