Improve model card: Add GitHub link and Python usage example
#1
by
nielsr
HF Staff
- opened
README.md
CHANGED
|
@@ -1,20 +1,20 @@
|
|
| 1 |
---
|
| 2 |
-
pipeline_tag: text-generation
|
| 3 |
library_name: transformers
|
| 4 |
license: cc-by-nc-4.0
|
|
|
|
| 5 |
tags:
|
| 6 |
- text-to-sql
|
| 7 |
- reinforcement-learning
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
| 11 |
# SLM-SQL: An Exploration of Small Language Models for Text-to-SQL
|
| 12 |
|
| 13 |
### Important Links
|
| 14 |
|
| 15 |
-
📖[Arxiv Paper](https://arxiv.org/abs/2507.22478) |
|
| 16 |
-
|
| 17 |
-
|
|
|
|
| 18 |
|
| 19 |
## News
|
| 20 |
|
|
@@ -35,8 +35,8 @@ tags:
|
|
| 35 |
> effectiveness
|
| 36 |
> and generalizability of our method, SLM-SQL. On the BIRD development set, the five evaluated models achieved an
|
| 37 |
> average
|
| 38 |
-
> improvement of 31.4 points. Notably, the 0.5B model reached 56.87
|
| 39 |
-
> achieved 67.08
|
| 40 |
|
| 41 |
### Framework
|
| 42 |
|
|
@@ -55,29 +55,89 @@ Performance Comparison of different Text-to-SQL methods on BIRD dev and test dat
|
|
| 55 |
|
| 56 |
<img src="https://raw.githubusercontent.com/CycloneBoy/slm_sql/main/data/image/slmsql_ablation_study.png" height="300" alt="slmsql_ablation_study">
|
| 57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
## Model
|
| 59 |
|
| 60 |
| **Model** | Base Model | Train Method | Modelscope | HuggingFace |
|
| 61 |
|------------------------------------------|------------------------------|--------------|---------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------|
|
| 62 |
-
| SLM-SQL-Base-0.5B | Qwen2.5-Coder-0.5B-Instruct | SFT | [
|
| 63 |
-
| SLM-SQL-0.5B | Qwen2.5-Coder-0.5B-Instruct | SFT + GRPO | [
|
| 64 |
-
| CscSQL-Merge-Qwen2.5-Coder-0.5B-Instruct | Qwen2.5-Coder-0.5B-Instruct | SFT + GRPO | [
|
| 65 |
-
| SLM-SQL-Base-1.5B | Qwen2.5-Coder-1.5B-Instruct | SFT | [
|
| 66 |
-
| SLM-SQL-1.5B | Qwen2.5-Coder-1.5B-Instruct | SFT + GRPO | [
|
| 67 |
-
| CscSQL-Merge-Qwen2.5-Coder-1.5B-Instruct | Qwen2.5-Coder-1.5B-Instruct | SFT + GRPO | [
|
| 68 |
-
| SLM-SQL-Base-0.6B | Qwen3-0.6B | SFT | [
|
| 69 |
-
| SLM-SQL-0.6B | Qwen3-0.6B | SFT + GRPO | [
|
| 70 |
-
| SLM-SQL-Base-1.3B | deepseek-coder-1.3b-instruct | SFT | [
|
| 71 |
-
| SLM-SQL-1.3B | deepseek-coder-1.3b-instruct | SFT + GRPO | [
|
| 72 |
-
| SLM-SQL-Base-1B | Llama-3.2-1B-Instruct | SFT | [
|
| 73 |
|
| 74 |
## Dataset
|
| 75 |
|
| 76 |
| **Dataset** | Modelscope | HuggingFace |
|
| 77 |
|----------------------------|------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------|
|
| 78 |
-
| SynsQL-Think-916k | [
|
| 79 |
-
| SynsQL-Merge-Think-310k | [
|
| 80 |
-
| bird train and dev dataset | [
|
| 81 |
|
| 82 |
## TODO
|
| 83 |
|
|
|
|
| 1 |
---
|
|
|
|
| 2 |
library_name: transformers
|
| 3 |
license: cc-by-nc-4.0
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
tags:
|
| 6 |
- text-to-sql
|
| 7 |
- reinforcement-learning
|
| 8 |
---
|
| 9 |
|
|
|
|
| 10 |
# SLM-SQL: An Exploration of Small Language Models for Text-to-SQL
|
| 11 |
|
| 12 |
### Important Links
|
| 13 |
|
| 14 |
+
📖[Arxiv Paper](https://arxiv.org/abs/2507.22478) |
|
| 15 |
+
\ud83d\udcbb[GitHub](https://github.com/CycloneBoy/slm_sql) |
|
| 16 |
+
\ud83e\udd17[HuggingFace](https://huggingface.co/collections/cycloneboy/slm-sql-688b02f99f958d7a417658dc) |
|
| 17 |
+
\ud83e\udd16[ModelScope](https://modelscope.cn/collections/SLM-SQL-624bb6a60e9643) |
|
| 18 |
|
| 19 |
## News
|
| 20 |
|
|
|
|
| 35 |
> effectiveness
|
| 36 |
> and generalizability of our method, SLM-SQL. On the BIRD development set, the five evaluated models achieved an
|
| 37 |
> average
|
| 38 |
+
> improvement of 31.4 points. Notably, the 0.5B model reached 56.87% execution accuracy (EX), while the 1.5B model
|
| 39 |
+
> achieved 67.08% EX. We will release our dataset, model, and code to github: https://github.com/CycloneBoy/slm_sql.
|
| 40 |
|
| 41 |
### Framework
|
| 42 |
|
|
|
|
| 55 |
|
| 56 |
<img src="https://raw.githubusercontent.com/CycloneBoy/slm_sql/main/data/image/slmsql_ablation_study.png" height="300" alt="slmsql_ablation_study">
|
| 57 |
|
| 58 |
+
## Usage
|
| 59 |
+
|
| 60 |
+
Here's how to use the model for Text-to-SQL generation.
|
| 61 |
+
|
| 62 |
+
```python
|
| 63 |
+
import torch
|
| 64 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 65 |
+
|
| 66 |
+
model_id = "cycloneboy/SLM-SQL-0.5B" # Or choose another model from the table above
|
| 67 |
+
|
| 68 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
| 69 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 70 |
+
model_id,
|
| 71 |
+
torch_dtype=torch.bfloat16,
|
| 72 |
+
device_map="auto",
|
| 73 |
+
)
|
| 74 |
+
|
| 75 |
+
# Example query
|
| 76 |
+
question = "How many members are there in the department of 'Sales'?"
|
| 77 |
+
|
| 78 |
+
# The chat template is critical for proper inference as the model is instruction-tuned.
|
| 79 |
+
messages = [
|
| 80 |
+
{"role": "system", "content": "You are an AI programming assistant, utilizing the Deepseek Coder model, developed by Deepseek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer
|
| 81 |
+
"},
|
| 82 |
+
{"role": "user", "content": f"### Instruction:
|
| 83 |
+
Generate a SQL query for the following question:
|
| 84 |
+
{question}
|
| 85 |
+
"},
|
| 86 |
+
]
|
| 87 |
+
|
| 88 |
+
# Apply the chat template to get the formatted prompt string
|
| 89 |
+
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 90 |
+
|
| 91 |
+
# Tokenize and generate
|
| 92 |
+
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)
|
| 93 |
+
|
| 94 |
+
outputs = model.generate(input_ids, max_new_tokens=256, do_sample=True, temperature=0.01, top_p=0.95)
|
| 95 |
+
|
| 96 |
+
# Decode the generated text, skipping special tokens
|
| 97 |
+
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 98 |
+
|
| 99 |
+
# Extract only the model's response based on the chat template's structure
|
| 100 |
+
# The response should start after "### Response:
|
| 101 |
+
" and end before "<|EOT|>
|
| 102 |
+
"
|
| 103 |
+
response_start = generated_text.find("### Response:
|
| 104 |
+
")
|
| 105 |
+
if response_start != -1:
|
| 106 |
+
response_content = generated_text[response_start + len("### Response:
|
| 107 |
+
"):]
|
| 108 |
+
response_end = response_content.find("<|EOT|>")
|
| 109 |
+
if response_end != -1:
|
| 110 |
+
sql_query = response_content[:response_end].strip()
|
| 111 |
+
print(f"Generated SQL: {sql_query}")
|
| 112 |
+
else:
|
| 113 |
+
print(f"Generated text (full): {response_content.strip()}")
|
| 114 |
+
else:
|
| 115 |
+
print(f"Generated text (full): {generated_text.strip()}")
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
## Model
|
| 119 |
|
| 120 |
| **Model** | Base Model | Train Method | Modelscope | HuggingFace |
|
| 121 |
|------------------------------------------|------------------------------|--------------|---------------------------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------|
|
| 122 |
+
| SLM-SQL-Base-0.5B | Qwen2.5-Coder-0.5B-Instruct | SFT | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-Base-0.5B) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-Base-0.5B) |
|
| 123 |
+
| SLM-SQL-0.5B | Qwen2.5-Coder-0.5B-Instruct | SFT + GRPO | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-0.5B) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-0.5B) |
|
| 124 |
+
| CscSQL-Merge-Qwen2.5-Coder-0.5B-Instruct | Qwen2.5-Coder-0.5B-Instruct | SFT + GRPO | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/CscSQL-Merge-Qwen2.5-Coder-0.5B-Instruct) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Merge-Qwen2.5-Coder-0.5B-Instruct) |
|
| 125 |
+
| SLM-SQL-Base-1.5B | Qwen2.5-Coder-1.5B-Instruct | SFT | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-Base-1.5B) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-Base-1.5B) |
|
| 126 |
+
| SLM-SQL-1.5B | Qwen2.5-Coder-1.5B-Instruct | SFT + GRPO | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-1.5B) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-1.5B) |
|
| 127 |
+
| CscSQL-Merge-Qwen2.5-Coder-1.5B-Instruct | Qwen2.5-Coder-1.5B-Instruct | SFT + GRPO | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/CscSQL-Merge-Qwen2.5-Coder-1.5B-Instruct) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/CscSQL-Merge-Qwen2.5-Coder-1.5B-Instruct) |
|
| 128 |
+
| SLM-SQL-Base-0.6B | Qwen3-0.6B | SFT | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-Base-0.6B) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-Base-0.6B) |
|
| 129 |
+
| SLM-SQL-0.6B | Qwen3-0.6B | SFT + GRPO | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-0.6B) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-0.6B) |
|
| 130 |
+
| SLM-SQL-Base-1.3B | deepseek-coder-1.3b-instruct | SFT | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-Base-1.3B ) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-Base-1.3B ) |
|
| 131 |
+
| SLM-SQL-1.3B | deepseek-coder-1.3b-instruct | SFT + GRPO | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-1.3B ) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-1.3B ) |
|
| 132 |
+
| SLM-SQL-Base-1B | Llama-3.2-1B-Instruct | SFT | [\ud83e\udd16 Modelscope](https://modelscope.cn/models/cycloneboy/SLM-SQL-Base-1B ) | [\ud83e\udd17 HuggingFace](https://huggingface.co/cycloneboy/SLM-SQL-Base-1B ) |
|
| 133 |
|
| 134 |
## Dataset
|
| 135 |
|
| 136 |
| **Dataset** | Modelscope | HuggingFace |
|
| 137 |
|----------------------------|------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------|
|
| 138 |
+
| SynsQL-Think-916k | [\ud83e\udd16 Modelscope](https://modelscope.cn/datasets/cycloneboy/SynsQL-Think-916k) | [\ud83e\udd17 HuggingFace](https://huggingface.co/datasets/cycloneboy/SynsQL-Think-916k) |
|
| 139 |
+
| SynsQL-Merge-Think-310k | [\ud83e\udd16 Modelscope](https://modelscope.cn/datasets/cycloneboy/SynsQL-Merge-Think-310k) | [\ud83e\udd17 HuggingFace](https://huggingface.co/datasets/cycloneboy/SynsQL-Merge-Think-310k) |
|
| 140 |
+
| bird train and dev dataset | [\ud83e\udd16 Modelscope](https://modelscope.cn/datasets/cycloneboy/bird_train) | [\ud83e\udd17 HuggingFace](https://huggingface.co/datasets/cycloneboy/bird_train) |
|
| 141 |
|
| 142 |
## TODO
|
| 143 |
|