README.md

---
license: apache-2.0
base_model: unsloth/Qwen3.5-4B
datasets:
  - Shumatsurontek/neo-sql-reasoning-combined
tags:
  - sql
  - text-to-sql
  - fine-tuned
  - lora
  - sft
  - trl
  - unsloth
  - neo-deep-agent-lab
  - modal
language:
  - en
pipeline_tag: text-generation
library_name: transformers
model-index:
  - name: Qwen3.5-4B-neo
    results:
      - task:
          type: text-generation
          name: Text-to-SQL
        dataset:
          name: Shumatsurontek/neo-sql-reasoning-combined
          type: Shumatsurontek/neo-sql-reasoning-combined
        metrics:
          - name: Training Loss
            type: loss
            value: 0.5880
---

<div align="center">

# Qwen3.5-4B-neo

**unsloth/Qwen3.5-4B** fine-tuned for **Text-to-SQL** generation via LoRA SFT

[![Built with neo-deep-agent-lab](https://img.shields.io/badge/built%20with-neo--deep--agent--lab-blue)](https://github.com/Shumatsurontek/neo-deep-agent-lab)
[![Modal](https://img.shields.io/badge/infra-Modal-purple)](https://modal.com)
[![Dataset](https://img.shields.io/badge/dataset-Shumatsurontek%2Fneo--sql--reasoning--combined-yellow)](https://huggingface.co/datasets/Shumatsurontek/neo-sql-reasoning-combined)

</div>

---

## Overview

This model was fine-tuned from [`unsloth/Qwen3.5-4B`](https://huggingface.co/unsloth/Qwen3.5-4B) on the
[`Shumatsurontek/neo-sql-reasoning-combined`](https://huggingface.co/datasets/Shumatsurontek/neo-sql-reasoning-combined) dataset using **Supervised Fine-Tuning (SFT)**
with **LoRA** adapters (merged into base weights for easy deployment).

- **Developed by:** [Shumatsurontek](https://github.com/Shumatsurontek)
- **Fine-tuning pipeline:** [neo-deep-agent-lab](https://github.com/Shumatsurontek/neo-deep-agent-lab)
- **Engine:** Unsloth
- **Compute:** NVIDIA L40S on [Modal](https://modal.com) serverless GPUs
- **Precision:** bf16
- **License:** Apache 2.0

## Training Details

### Configuration

| Parameter | Value |
|---|---|
| Base model | `unsloth/Qwen3.5-4B` |
| Method | SFT + LoRA (bf16, merged) |
| Learning rate | 2e-04 (cosine schedule, 5% warmup) |
| LoRA rank (r) | 16 |
| LoRA alpha | 32 (α/r = 32/16) |
| Batch size | 4 per GPU |
| Max sequence length | 2048 |
| Epochs | 2 |

### Hyperparameters

$$
\begin{aligned}
\eta &= 2 \times 10^{-4} & \text{(learning rate, cosine schedule)} \\
r &= 16 & \text{(LoRA rank)} \\
\alpha &= 32 & \text{(LoRA scaling, } \alpha/r = 2 \text{)} \\
B &= 4 & \text{(batch size per GPU)} \\
T &= 2048 & \text{(max sequence length)} \\
E &= 2 & \text{(epochs)} \\
\end{aligned}
$$

### Results

| Metric | Value |
|---|---|
| **Final training loss** | **0.5880** |
| Total optimization steps | 1,914 |
| Warmup | 5% (linear) |
| LR scheduler | cosine → 0 |
| Optimizer | AdamW 8-bit |

## Dataset

[`Shumatsurontek/neo-sql-reasoning-combined`](https://huggingface.co/datasets/Shumatsurontek/neo-sql-reasoning-combined)

Each sample follows a 3-turn chat format:

```
System: You are a SQL expert. Given a database schema and a
        natural language question, generate the correct SQL query.
User:   Schema: CREATE TABLE orders (id INT, total DECIMAL);
        Question: What is the total revenue?
Assistant: SELECT SUM(total) FROM orders;
```

## Quickstart

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Shumatsurontek/Qwen3.5-4B-neo")
tokenizer = AutoTokenizer.from_pretrained("Shumatsurontek/Qwen3.5-4B-neo")

messages = [
    {"role": "system", "content": "You are a SQL expert. Given a database schema and a natural language question, generate the correct SQL query."},
    {"role": "user", "content": "Schema: CREATE TABLE orders (id INT, user_id INT, total DECIMAL);\nQuestion: Total revenue per user?"},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

### vLLM (OpenAI-compatible server)

```bash
vllm serve Shumatsurontek/Qwen3.5-4B-neo --trust-remote-code
```

```bash
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "Shumatsurontek/Qwen3.5-4B-neo",
  "messages": [
    {"role": "system", "content": "You are a SQL expert."},
    {"role": "user", "content": "Schema: CREATE TABLE users (id INT, name TEXT);\nQuestion: List all users?"}
  ]
}'
```

## Intended Use

This model is designed for **text-to-SQL** tasks: given a database schema and a natural-language question,
it generates the corresponding SQL query. Best suited for analytical and read-only queries.

**Out of scope:** DDL/DML generation (CREATE, DROP, INSERT, UPDATE, DELETE), multi-database queries,
or production use without human review of generated SQL.



## Benchmark Results

Evaluated against baseline [`unsloth/Qwen3.5-4B`](https://huggingface.co/unsloth/Qwen3.5-4B) using [lm-eval-harness](https://github.com/EleutherAI/lm-evaluation-harness) on NVIDIA L40S.

> Evaluated on 50 samples per task.

| Benchmark | Baseline | Finetuned | Delta |
|---|:---:|:---:|:---:|
| **MMLU: STEM** | 73.5 | 71.8 | 🔴 -1.7 |
| **MMLU: HUMANITIES** | 76.5 | 73.4 | 🔴 -3.1 |
| **HELLASWAG** | 48.0 | 52.0 | 🟢 +4.0 |
| **MMLU** | 77.1 | 74.7 | 🔴 -2.4 |
| **MMLU: OTHER** | 77.7 | 75.8 | 🔴 -1.8 |
| **ARC_CHALLENGE** | 60.0 | 60.0 | ⚪ 0.0 |
| **MMLU: SOCIAL SCIENCES** | 83.0 | 79.7 | 🔴 -3.3 |

## Citation

```bibtex
@misc{Qwen3_5_4B_neo,
  title  = {Qwen3.5-4B-neo},
  author = {Shumatsurontek},
  year   = {2026},
  url    = {https://huggingface.co/Shumatsurontek/Qwen3.5-4B-neo}
}
```

## License

Apache 2.0