File size: 4,504 Bytes
0440e12
 
 
 
 
 
 
 
d6b8247
 
 
 
 
 
 
 
 
 
 
 
 
 
a88aee5
d6b8247
 
 
 
 
 
 
 
 
7e2fb51
d6b8247
 
 
 
 
 
 
 
 
 
a88aee5
 
d6b8247
a88aee5
 
d6b8247
 
 
4ccba91
d6b8247
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: other
language:
- en
pipeline_tag: text-generation
tags:
- code
---
# Defog SQLCoder
Defog's SQLCoder is a state-of-the-art LLM for converting natural language questions to SQL queries.

[Interactive Demo](https://defog.ai/sqlcoder-demo/) | [🤗 HF Repo](https://huggingface.co/defog/sqlcoder2) | [♾️ Colab](https://colab.research.google.com/drive/1z4rmOEiFkxkMiecAWeTUlPl0OmKgfEu7?usp=sharing) | [🐦 Twitter](https://twitter.com/defogdata)

## TL;DR
SQLCoder is a 15B parameter model that outperforms `gpt-3.5-turbo` for natural language to SQL generation tasks on our [sql-eval](https://github.com/defog-ai/sql-eval) framework, and significantly outperforms all popular open-source models. When fine-tuned on a given schema, it also outperforms `gpt-4`

SQLCoder is fine-tuned on a base StarCoder model.

## Results on novel datasets not seen in training
| model   | perc_correct |
|-|-|  
| gpt4-2023-10-04    | 82.0 |
| defog-sqlcoder2    | 77.5 |
| gpt4-2023-08-28    | 74.0 |
| defog-sqlcoder-7b  | 71.0 |
| gpt-3.5-2023-10-04 | 66.0 |
| claude-2           | 64.5 |
| gpt-3.5-2023-08-28 | 61.0 |
| claude_instant_1   | 61.0 |
| text-davinci-003   | 52.5 |

## License
The code in this repo (what little there is of it) is Apache-2 licensed. The model weights have a `CC BY-SA 4.0` license, with additional responsible use restrictions added. The TL;DR is that you can use and modify the model for any purpose – including commercial use. However, if you modify the weights (for example, by fine-tuning), you must open-source your modified weights under the same license terms.

## Training
Defog was trained on more than 20,000 human-curated questions. These questions were based on 10 different schemas. None of the schemas in the training data were included in our evaluation framework. 

You can read more about our [training approach](https://defog.ai/blog/open-sourcing-sqlcoder2-7b/) and [evaluation framework](https://defog.ai/blog/open-sourcing-sqleval/).

## Results by question category
We classified each generated question into one of 5 categories. The table displays the percentage of questions answered correctly by each model, broken down by category.
| query_category   |   gpt-4 |   sqlcoder2-15b |   sqlcoder-7b |   gpt-3.5 |   claude-2 |   claude-instant |   gpt-3 |
|:-----------------|--------:|----------------:|--------------:|----------:|-----------:|-----------------:|--------:|
| date             |    72   |            80   |          64   |      68   |       52   |             48   |    32   |
| group_by         |    91.4 |            82.9 |          82.9 |      77.1 |       71.4 |             71.4 |    71.4 |
| order_by         |    82.9 |            77.1 |          74.3 |      68.6 |       74.3 |             74.3 |    68.6 |
| ratio            |    80   |            74.3 |          54.3 |      37.1 |       57.1 |             45.7 |    25.7 |
| join             |    82.9 |            74.3 |          74.3 |      71.4 |       65.7 |             62.9 |    57.1 |
| where            |    80   |            77.1 |          74.3 |      74.3 |       62.9 |             60   |    54.3 |

## Using SQLCoder
You can use SQLCoder via the `transformers` library by downloading our model weights from the Hugging Face repo. We have added sample code for [inference](https://github.com/defog-ai/sqlcoder/blob/main/inference.py) on a [sample database schema](https://github.com/defog-ai/sqlcoder/blob/main/metadata.sql). 
```bash
python inference.py -q "Question about the sample database goes here"

# Sample question:
# Do we get more revenue from customers in New York compared to customers in San Francisco? Give me the total revenue for each city, and the difference between the two.
```

You can also use a demo on our website [here](https://defog.ai/sqlcoder-demo), or run SQLCoder in Colab [here](https://colab.research.google.com/drive/13BIKsqHnPOBcQ-ba2p77L5saiepTIwu0#scrollTo=ZpbVgVHMkJvC)

## Hardware Requirements
SQLCoder has been tested on an A100 40GB GPU with `bfloat16` weights. You can also load an 8-bit and 4-bit quantized version of the model on consumer GPUs with 20GB or more of memory – like RTX 4090, RTX 3090, and Apple M2 Pro, M2 Max, or M2 Ultra Chips with 20GB or more of memory.

## Todo

- [x] Open-source the v1 model weights
- [x] Train the model on more data, with higher data variance
- [ ] Tune the model further with Reward Modelling and RLHF
- [ ] Pretrain a model from scratch that specializes in SQL analysis