File size: 4,087 Bytes
eaecfc9
 
21bbd2e
 
eaecfc9
21bbd2e
 
 
 
 
 
eaecfc9
 
21bbd2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
eaecfc9
21bbd2e
eaecfc9
21bbd2e
eaecfc9
21bbd2e
 
 
eaecfc9
21bbd2e
eaecfc9
21bbd2e
 
 
eaecfc9
21bbd2e
eaecfc9
21bbd2e
 
eaecfc9
21bbd2e
eaecfc9
21bbd2e
 
 
 
 
 
 
eaecfc9
21bbd2e
eaecfc9
21bbd2e
 
 
eaecfc9
21bbd2e
 
 
eaecfc9
21bbd2e
 
 
 
 
 
eaecfc9
21bbd2e
 
 
 
eaecfc9
21bbd2e
eaecfc9
21bbd2e
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
---
base_model: Qwen/Qwen3-4B-Base
library_name: transformers
pipeline_tag: text-generation
tags:
  - text-to-sql
  - sql
  - qwen3
  - llamafactory
  - spider
  - spider-test-suite
---

# qwen_4b_sql

`qwen_4b_sql` is a `Qwen3-4B-Base` model finetuned for text-to-SQL generation with full SFT on a cleaned split of `PipableAI/pip-txt-to-sql-spider-bird-dataset`.

This repository tracks the stronger 4B checkpoint from our H20 single-GPU training runs. In our internal comparison, this checkpoint outperformed the corresponding `Qwen3-1.7B-Base` baseline on Spider execution accuracy.

## Base Model

- Base model: [`Qwen/Qwen3-4B-Base`](https://huggingface.co/Qwen/Qwen3-4B-Base)
- Finetuning framework: `LLaMA-Factory`
- Training mode: `Full SFT`
- Task: `schema + question -> SQL only`

## Training Data

- Primary dataset: [`PipableAI/pip-txt-to-sql-spider-bird-dataset`](https://huggingface.co/datasets/PipableAI/pip-txt-to-sql-spider-bird-dataset)
- We used a cleaned local split derived from that dataset for train/validation

## Training Setup

- Hardware: single `NVIDIA H20 96GB`
- Precision: `bf16`
- Context length: `2048`
- Per-device train batch size: `1`
- Gradient accumulation steps: `8`
- Effective batch size: `8`
- Learning rate: `5e-6`
- Scheduler: `cosine`
- Warmup steps: `300`
- Epochs: `4.0`
- Template: `qwen3_nothink`
- Best-checkpoint selection: `load_best_model_at_end = true`

## Spider Benchmark

The following numbers are from Spider dev using the official evaluation tooling:

- Official `match` evaluation from `test-suite-sql-eval`
- Official Spider `Test Suite` execution evaluation

### Main Results

| Metric | Score |
| --- | ---: |
| Spider official exact match | 35.0% |
| Spider Test Suite execution accuracy | 67.6% |

### Difficulty Breakdown

| Difficulty | Exact Match | Test Suite Exec |
| --- | ---: | ---: |
| Easy | 64.9% | 87.5% |
| Medium | 37.4% | 72.9% |
| Hard | 16.1% | 50.0% |
| Extra | 3.6% | 42.2% |

## Notes

- This model is stronger under execution-based Spider evaluation than our best `Qwen3-1.7B-Base` run.
- In our experiments, exact-match metrics were often stricter than execution-based metrics because semantically valid SQL rewrites do not always match the Spider gold form exactly.
- A later 4B rerun with altered training settings underperformed this checkpoint on Spider and is not the checkpoint published here.

## Intended Use

This model is intended for:

- text-to-SQL research baselines
- schema-conditioned SQL generation experiments
- single-turn SQL generation from natural language plus schema text

It is not validated for:

- production-grade database access control
- unrestricted execution over arbitrary enterprise schemas
- multi-turn agent workflows without extra prompting / tooling

## Example Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "bsq1989/qwen_4b_sql"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

prompt = """Generate SQL from the given schema and question. Output SQL only.

Schema:
CREATE TABLE twitter (TweetID INTEGER, UserID INTEGER, LocationID INTEGER, Lang TEXT, ...);
CREATE TABLE location (LocationID INTEGER, Country TEXT, City TEXT, ...);

Question:
How many tweets are in English?
"""

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Limitations

- Performance drops on more open-ended and heterogeneous SQL benchmarks than Spider.
- The model can still produce invalid column references on out-of-distribution schemas.
- Benchmark numbers here reflect our current internal setup and should be reproduced with the same evaluation pipeline for strict comparison.