Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,123 @@
|
|
1 |
---
|
|
|
|
|
2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
license: apache-2.0
|
5 |
+
tags:
|
6 |
+
- zsql
|
7 |
+
- chatml
|
8 |
+
- synthetic data
|
9 |
+
- text-to-sql
|
10 |
+
- dpo
|
11 |
+
datasets:
|
12 |
+
- zerolink/zsql-postgres-dpo
|
13 |
+
widget:
|
14 |
+
- text: >-
|
15 |
+
<|im_start|>system
|
16 |
+
Translate English to Postgres SQL.<|im_end|>
|
17 |
+
<|im_start|>user
|
18 |
+
Using the schema:
|
19 |
+
CREATE TABLE Product (
|
20 |
+
product_id INTEGER PRIMARY KEY,
|
21 |
+
name TEXT NOT NULL,
|
22 |
+
price DECIMAL NOT NULL,
|
23 |
+
description TEXT
|
24 |
+
);
|
25 |
+
Generate SQL for the following question:
|
26 |
+
What are all products worth more than $5.10?
|
27 |
+
<|im_end|>
|
28 |
+
example_title: sql
|
29 |
---
|
30 |
+
|
31 |
+
zsql-postgres is a text-to-SQL model which is instruction tuned for SQL query
|
32 |
+
synthesis on English language text to the Postgres SQL code. The model is trained
|
33 |
+
on the [ZeroLink DPO](https://huggingface.co/datasets/zerolink/zsql-postgres-dpo)
|
34 |
+
dataset.
|
35 |
+
|
36 |
+
This model is only capable of generating SQL queries and is designed to be
|
37 |
+
further fine-tuned to specific database schemas.
|
38 |
+
|
39 |
+
## Usage
|
40 |
+
|
41 |
+
You can run this model using the following code:
|
42 |
+
|
43 |
+
```python
|
44 |
+
import transformers
|
45 |
+
from transformers import AutoTokenizer
|
46 |
+
|
47 |
+
model = "zerolink/zsql-en-postgres"
|
48 |
+
|
49 |
+
tokenizer = AutoTokenizer.from_pretrained(model)
|
50 |
+
|
51 |
+
prompt = f"""
|
52 |
+
Using the schema:
|
53 |
+
CREATE TABLE Product (
|
54 |
+
product_id INTEGER PRIMARY KEY,
|
55 |
+
name TEXT NOT NULL,
|
56 |
+
price DECIMAL NOT NULL,
|
57 |
+
description TEXT
|
58 |
+
);
|
59 |
+
|
60 |
+
CREATE TABLE Customer (
|
61 |
+
customer_id INTEGER PRIMARY KEY,
|
62 |
+
name TEXT NOT NULL,
|
63 |
+
email TEXT,
|
64 |
+
phone TEXT
|
65 |
+
);
|
66 |
+
Generate SQL for the following question:
|
67 |
+
What are the prices and descriptions for all products that are greater than $5?
|
68 |
+
"""
|
69 |
+
|
70 |
+
system = "Translate English to Postgres SQL."
|
71 |
+
message = [
|
72 |
+
{"role": "system", "content": system},
|
73 |
+
{"role": "user", "content": prompt},
|
74 |
+
]
|
75 |
+
|
76 |
+
prompt = tokenizer.apply_chat_template(message, add_generation_prompt=True, tokenize=False)
|
77 |
+
|
78 |
+
# Create pipeline
|
79 |
+
pipeline = transformers.pipeline(
|
80 |
+
"text-generation",
|
81 |
+
model=model,
|
82 |
+
tokenizer=tokenizer
|
83 |
+
)
|
84 |
+
|
85 |
+
# Generate text
|
86 |
+
sequences = pipeline(
|
87 |
+
prompt,
|
88 |
+
do_sample=True,
|
89 |
+
temperature=0.1,
|
90 |
+
top_p=0.9,
|
91 |
+
num_return_sequences=1,
|
92 |
+
max_length=1024,
|
93 |
+
)
|
94 |
+
print(sequences[0]['generated_text'])
|
95 |
+
```
|
96 |
+
|
97 |
+
## Training hyperparameters
|
98 |
+
|
99 |
+
**LoRA**:
|
100 |
+
|
101 |
+
* r=16
|
102 |
+
* lora_alpha=16
|
103 |
+
* lora_dropout=0.05
|
104 |
+
* bias="none"
|
105 |
+
* task_type="CAUSAL_LM"
|
106 |
+
* target_modules=['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']
|
107 |
+
|
108 |
+
**Training arguments**:
|
109 |
+
|
110 |
+
* per_device_train_batch_size=4
|
111 |
+
* gradient_accumulation_steps=4
|
112 |
+
* gradient_checkpointing=True
|
113 |
+
* learning_rate=5e-5
|
114 |
+
* lr_scheduler_type="linear"
|
115 |
+
* max_steps=200
|
116 |
+
* optim="paged_adamw_32bit"
|
117 |
+
* warmup_steps=100
|
118 |
+
|
119 |
+
**DPOTrainer**:
|
120 |
+
|
121 |
+
* beta=0.1
|
122 |
+
* max_prompt_length=4096
|
123 |
+
* max_length=3516
|