|
|
--- |
|
|
tags: |
|
|
- text-to-sql |
|
|
- qwen |
|
|
- tencent-trac3 |
|
|
- fine-tuned |
|
|
license: apache-2.0 |
|
|
--- |
|
|
|
|
|
# wexhi/trac3_sql |
|
|
|
|
|
## 模型描述 |
|
|
|
|
|
这是一个基于 **Qwen** 微调的**全量模型**，专门用于 SQL 生成任务（Text-to-SQL）。 |
|
|
|
|
|
训练数据来自 Tencent TRAC3 数据集，采用**记忆化训练策略**，目标是在训练集上达到 100% 准确率。 |
|
|
|
|
|
## 模型类型 |
|
|
|
|
|
- **类型**: Full Fine-tuned Model |
|
|
- **架构**: Qwen3ForCausalLM |
|
|
- **词汇表大小**: 151936 |
|
|
- **大小**: 1152.06 MB |
|
|
|
|
|
## 使用方法 |
|
|
|
|
|
### 1. 安装依赖 |
|
|
|
|
|
```bash |
|
|
pip install transformers torch |
|
|
``` |
|
|
|
|
|
### 2. 加载模型 |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
|
"wexhi/trac3_sql", |
|
|
torch_dtype="auto", |
|
|
device_map="auto", |
|
|
trust_remote_code=True, |
|
|
) |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained( |
|
|
"wexhi/trac3_sql", |
|
|
trust_remote_code=True, |
|
|
) |
|
|
``` |
|
|
|
|
|
### 3. 生成 SQL |
|
|
|
|
|
```python |
|
|
messages = [ |
|
|
{"role": "system", "content": "You are a SQL generator. Generate SQL in this format:\n```sql\n...\n```"}, |
|
|
{"role": "user", "content": "ID: 1\n\nQuestion:\nWhat is the total revenue?"} |
|
|
] |
|
|
|
|
|
prompt = tokenizer.apply_chat_template( |
|
|
messages, |
|
|
tokenize=False, |
|
|
add_generation_prompt=True, |
|
|
enable_thinking=False, |
|
|
) |
|
|
|
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
|
|
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.0) |
|
|
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True) |
|
|
print(response) |
|
|
``` |
|
|
|
|
|
### 4. 使用 vLLM 加速（推荐） |
|
|
|
|
|
```bash |
|
|
pip install vllm |
|
|
``` |
|
|
|
|
|
```python |
|
|
from vllm import LLM, SamplingParams |
|
|
|
|
|
llm = LLM(model="wexhi/trac3_sql", trust_remote_code=True) |
|
|
sampling_params = SamplingParams(temperature=0.0, max_tokens=512) |
|
|
|
|
|
prompts = [...] # 批量 prompts |
|
|
outputs = llm.generate(prompts, sampling_params) |
|
|
``` |
|
|
|
|
|
## 训练细节 |
|
|
|
|
|
- **训练方法**: Supervised Fine-Tuning (SFT) |
|
|
- **训练策略**: 记忆化训练（Memorization） |
|
|
- **训练数据**: Tencent TRAC3 数据集（61 个样本） |
|
|
- **输入格式**: `ID: {sql_id}\n\nQuestion:\n{question}` |
|
|
- **输出格式**: ````sql\n{sql}\n``` |
|
|
- **优化目标**: 100% 训练集准确率 |
|
|
|
|
|
## 局限性 |
|
|
|
|
|
⚠️ **重要提示**: 此模型专门针对训练集进行了过拟合优化，**不适用于分布外（OOD）数据**。 |
|
|
|
|
|
- ✅ 对于训练集中的问题，能够准确生成 SQL |
|
|
- ❌ 对于未见过的问题，可能无法正确泛化 |
|
|
|
|
|
## License |
|
|
|
|
|
Apache 2.0 |
|
|
|
|
|
## 引用 |
|
|
|
|
|
如果使用了此模型，请引用： |
|
|
|
|
|
``` |
|
|
Tencent TRAC3 Challenge - Text-to-SQL Fine-tuned Model |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
*Created: 2025-11-24* |
|
|
|
