devansh173/orashift-qlora-adapter

OraShift: Oracle SQL to PostgreSQL (QLoRA adapter)

A LoRA adapter for Qwen2.5-Coder-3B-Instruct that translates Oracle SQL into PostgreSQL. Every number on this card is execution accuracy: a translation counts as correct only if it runs on PostgreSQL 17 and returns what the original statement returns on Oracle Database 26ai. Text similarity is not used anywhere.

• Code, evaluation and full write-up: https://github.com/devansh173/orashift
• Demo: https://huggingface.co/spaces/devansh173/orashift
• Training data: https://huggingface.co/datasets/devansh173/orashift-pairs

Results

330 held-out test units, each scored by executing it on both engines.

Strategy	Execution accuracy
Base model, no adapter	245/330 (74%)
This adapter	285/330 (86%)
sqlglot (rule-based)	202/330 (61%)
Hybrid: sqlglot when it verifies, this adapter otherwise	288/330 (87%)

The test set holds out a whole schema and 25 of 123 templates, so generalisation is measured on two axes:

Split	Base	This adapter
Seen schema, seen template	26/33 (79%)	32/33 (97%)
Unseen schema, seen template	125/160 (78%)	158/160 (99%)
Seen schema, unseen template	69/102 (68%)	71/102 (70%)
Unseen schema, unseen template	25/35 (71%)	24/35 (69%)

The adapter generalises to unfamiliar table and column names, not to unfamiliar transformations. On the hardest split it is no better than the base model. On the unseen SYSDATE window template it often returns the Oracle input unchanged, scoring 2/26 on SYSDATE against the base model's 23/26.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-3B-Instruct")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-3B-Instruct", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "devansh173/orashift-qlora-adapter")

messages = [
    {
        "role": "system",
        "content": "You translate Oracle SQL into PostgreSQL. Return only the PostgreSQL statement.",
    },
    {"role": "user", "content": "SELECT employee_id FROM hr_employees WHERE ROWNUM <= 5"},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
output = model.generate(inputs, max_new_tokens=448, do_sample=False)
print(tokenizer.decode(output[0][inputs.shape[1] :], skip_special_tokens=True))

The adapter was trained with a longer system prompt that includes the task rules and the Oracle DDL of the tables a statement uses. For best results use the same format: see app/translate.py and the dataset's system turns.

Training

Setting	Value
Base	unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit (QLoRA, 4-bit)
LoRA	r=16, alpha=16, dropout 0, all 7 attention and MLP projections
Trainable parameters	29.9 M, about 1% of the 3.1 B base
Data	418 training pairs, every one execution-verified
Loss	Assistant tokens only
Schedule	3 epochs, lr 2e-4, linear, 5% warmup, effective batch 8, AdamW 8-bit
Sequence length	1024 (longest example 804 tokens)
Hardware	One Kaggle T4, fp16, 13.1 minutes
Final eval loss	0.0007

Limitations

• Does not generalise to unseen constructs; see the split table above.
• Synthetic, template-generated data: phrasing is more uniform than real-world Oracle SQL, so expect lower accuracy on hand-written production queries.
• Small training set (418 pairs). DDL and DML are thinly covered.
• Always verify output before use. In the project, every translation is checked by running it against both databases; this adapter alone does no such check.

Licence

The adapter is a derivative of Qwen2.5-Coder-3B-Instruct and is subject to the Qwen Research License, which does not permit commercial use. The OraShift code is MIT-licensed.