---
base_model:
- griffith-bigdata/Qwen-2.5-Coder-0.5B-SQL-Writer
license: apache-2.0
language:
- en
tags:
- text-to-sql
- spider
- grpo
- finer-sql
- code
library_name: transformers
pipeline_tag: text-generation
---

# FINER-SQL-0.5B-Spider

A small but capable 0.5 B-parameter Text-to-SQL model fine-tuned from
[`griffith-bigdata/Qwen-2.5-Coder-0.5B-SQL-Writer`](https://huggingface.co/griffith-bigdata/Qwen-2.5-Coder-0.5B-SQL-Writer)
with GRPO + the FINER-SQL dense rewards (Memory + Atomic).

✅ **75.0% Execution Accuracy on Spider Dev** (n=30, value-aware voting). Runs on a 4-8 GB GPU.

📄 See other models: https://huggingface.co/collections/griffith-bigdata/finer-sql
📄 GitHub: https://github.com/thanhdath/finer-sql/tree/main

---

## FINER-SQL Model Family — Comparison Across All Sizes

| Model | Params | BIRD Dev (n=30, vav) | Spider Dev (n=30, vav, +agg_hint) |
|-------|--------|---------------------|----------------------------------|
| [FINER-SQL-3B-BIRD](https://huggingface.co/griffith-bigdata/FINER-SQL-3B-BIRD) | 3 B | **67.54%** ✅ | 83.8% |
| [FINER-SQL-3B-Spider](https://huggingface.co/griffith-bigdata/FINER-SQL-3B-Spider) | 3 B | 63.04% | **85.10%** ✅ |
| [FINER-SQL-0.5B-BIRD](https://huggingface.co/griffith-bigdata/FINER-SQL-0.5B-BIRD) | 0.5 B | **50.85%** ✅ | 68.6% |
| **FINER-SQL-0.5B-Spider** *(this model)* | 0.5 B | TBD | **75.0%** ✅ |

The 0.5 B Spider model is **6.4 pp better** than the 0.5 B BIRD model on Spider Dev — confirming dataset-specific specialisation matters even at small scales.

---

## Inference

### Quick start (vLLM)

```python
from vllm import LLM, SamplingParams

llm = LLM(
    model="griffith-bigdata/FINER-SQL-0.5B-Spider",
    dtype="bfloat16",
    max_model_len=4096,
    gpu_memory_utilization=0.7,
)

system_prompt = """You are a meticulous SQL expert. Generate a single, correct SQL query for the user question and the provided database schema.
Follow this exact response format:

Rules:
- Output exactly one SQL statement.
- The SQL must be executable on SQLite.
- Do not include any explanatory text.
- Output one SQL statement only. Do not include any extra text, tags, or code fences."""

sampling = SamplingParams(n=30, temperature=1.0, max_tokens=2048)
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": f"Database Schema:\n{schema}\n\nQuestion: {question}"},
]
output = llm.chat(messages, sampling)
candidate_sqls = [c.text.split("</think>")[-1].strip() for c in output[0].outputs]
# Apply majority voting (vav) — see GitHub repo
```

### Recommended evaluation pipeline

1. Generate n=30 candidates with temperature=1.0
2. Execute each candidate; group results
3. Pick from the largest non-empty success group (value-aware voting, "vav")
4. Score with the official Spider evaluator (`test_suite_sql_eval`)

This pipeline gives **75.0% Spider Dev EX** (75.44% MV).

---

## Detailed Spider Dev results (n=30, vav)

| Hardness | Count | Execution Accuracy |
|----------|-------|--------------------|
| Easy | 248 | 91.9% |
| Medium | 446 | 82.5% |
| Hard | 174 | 62.6% |
| Extra Hard | 166 | 42.8% |
| **All** | **1034** | **75.0%** |

Recall@30: **85.11%** (any-correct rate among 30 candidates).

---

## Training

| Parameter | Value |
|-----------|-------|
| Base model | `griffith-bigdata/Qwen-2.5-Coder-0.5B-SQL-Writer` |
| Algorithm | GRPO |
| Train data | Spider train (8,659 samples) |
| Total steps | 2000 (this checkpoint = 2000) |
| Learning rate | 8e-6 |
| Num generations per prompt | 32 |
| Gradient accumulation | 32 |
| Max completion length | 2048 |
| Max prompt length | 1500 |
| Temperature (rollout) | 1.0 |
| Selection during eval | vav (value-aware voting) |
| Rewards | Execution + Atomic + Memory + Format |

---

## License

Inherits the base model's license (Apache 2.0).

---

## Citation

```bibtex
@article{finer-sql-2026,
  title  = {FINER-SQL: Fine-grained reasoning rewards for small Text-to-SQL models},
  author = {Thanh Dat and others},
  year   = {2026},
}
```