---
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
library_name: peft
license: apache-2.0
pipeline_tag: text-generation
language: [en]
datasets: [samrat-kar/ap-sql-peft]
tags:
- base_model:adapter:Qwen/Qwen2.5-Coder-7B-Instruct
- lora
- qlora
- text-to-sql
- oracle
- accounts-payable
---

# ap-sql-v1

A LoRA adapter for `Qwen/Qwen2.5-Coder-7B-Instruct` that turns Accounts Payable questions into read-only Oracle SQL.
It is built for an AI data-analyst assistant that answers questions over a schema modelled on Oracle Fusion AP,
Payments and Supplier tables.

The model gets a system prompt containing the relevant table DDL (from schema retrieval) and a business glossary.
It answers with a single ```` ```sql ```` block. It has also learned to fix a query when it is given the Oracle error
from its previous attempt, and to answer follow-up questions in a conversation.

## Results

Execution accuracy on the 179-case held-out test split. The generated SQL and the gold SQL were both run on Oracle 23ai
and their result sets compared. A case counts only when the rows match exactly. The adapter was served by
vLLM 0.12.0 on the AWQ base `Qwen/Qwen2.5-Coder-7B-Instruct-AWQ`.

| Model | T1 | T2 | T3 | T4 | T5 | Overall |
|---|---|---|---|---|---|---|
| Base `Qwen2.5-Coder-7B-Instruct-AWQ` | 57% | 32% | 21% | 32% | 56% | 35% |
| **ap-sql-v1** | **97%** | **96%** | **90%** | **89%** | **89%** | **93%** |

Invented-column errors (ORA-00904) fell from 43 cases to 0. The adapter adds about 1 second of median latency
(2.8 s to 3.8 s on an RTX 5070 Ti Laptop GPU).

**Caveats.** 149 of the 179 test cases are templated, as is most of the training set. No test case shares a template
group or an identical question with the training data, but real user questions will vary more than the test set.
The eval uses each case's stored prompt, so it measures the model on its own; schema retrieval is outside its scope.

## Usage

### vLLM

```bash
vllm serve Qwen/Qwen2.5-Coder-7B-Instruct-AWQ --quantization awq_marlin \
  --enable-lora --max-lora-rank 16 --lora-modules ap-sql-v1=samrat-kar/ap-sql-v1
```

Then send `model="ap-sql-v1"` to the OpenAI-compatible `/v1/chat/completions` endpoint.

### Transformers + PEFT

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "samrat-kar/ap-sql-v1")
tok = AutoTokenizer.from_pretrained("samrat-kar/ap-sql-v1")
```

Use the same system-prompt format as the training data (see the dataset). The schema must be given as `CREATE TABLE` DDL.

## Training

| | |
|---|---|
| Method | QLoRA: 4-bit NF4 base with double quantisation, bf16 compute |
| LoRA | r=16, alpha=32, dropout=0.05; q/k/v/o/gate/up/down projections |
| Data | 1,441 train / 95 val rows from `samrat-kar/ap-sql-peft` |
| Schedule | 3 epochs, 273 optimizer steps, lr 2e-4 cosine, effective batch 16, paged AdamW 8-bit |
| Max length | 3,840 tokens |
| Final train loss | 0.0157 |
| Compute | 2.5 h, peak 17.5 GB GPU memory |

The loss is computed on the assistant turn only.

## Limitations

- Trained on a single demo schema. Other schemas, or other Fusion modules, need new training data.
- Produces Oracle dialect only (`FETCH FIRST N ROWS ONLY`, `ADD_MONTHS`, `TRUNC`).
- Always run the output through read-only guardrails and a read-only database user. The model is not a security boundary.