File size: 3,936 Bytes
4acbe3a
 
 
 
 
 
 
 
b6a8eb3
216215c
 
 
 
 
b6a8eb3
216215c
c0f3ff6
 
216215c
 
 
cbd34b0
1a06498
216215c
 
 
 
 
 
 
 
 
 
 
5f5bfaa
216215c
75fdc0d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
216215c
 
 
 
 
 
 
 
 
b6a8eb3
 
65039e8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4acbe3a
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
---
license: llama2
datasets:
- bigcode/the-stack
- NumbersStation/NSText2SQL
language:
- en
---
# SambaCoder-nsql-llama-2-70b

## Model Description

NSQL is a family of autoregressive open-source large foundation models (FMs) designed specifically for SQL generation tasks.

In this repository we are introducing a new member of NSQL, SambaCoder-nsql-llama-2-70b. It's based on Meta's original [Llama-2 70B model](https://huggingface.co/meta-llama/Llama-2-70b) and further pre-trained on a dataset of general SQL queries and then fine-tuned on a dataset composed of text-to-SQL pairs.

Use of this model is governed by the Meta’s Llama 2 Community License Agreement. Please review and accept the license before downloading the model weights and tokenizer

### Basic Information

<!-- Provide the basic links for the model. -->
- **Blog Post**: [Link](https://sambanova.ai/blog/sambacoder-nsql-llama-2-70b-model)
- **HF Hosting**: [Chat with me!](https://huggingface.co/spaces/sambanovasystems/nova-nsql-Llama-2-70B-text-to-SQL)

## Training Data

The general SQL queries are the SQL subset from [The Stack](https://huggingface.co/datasets/bigcode/the-stack), containing 1M training samples. The labeled text-to-SQL pairs come from the NSText2SQL dataset (https://huggingface.co/datasets/NumbersStation/NSText2SQL).

## Evaluation Data

We evaluate our models on three text-to-SQL benchmarks: Spider, Bird, and text2sql.

## Training Procedure

SambaCoder-nsql-llama-2-70b was trained using cross-entropy loss to maximize the likelihood of sequential inputs. For finetuning on text-to-SQL pairs, we only compute the loss over the SQL portion of the pair. The model is trained using SambaNova's in-house Reconfigurable Dataflow Unit (RDU), leveraging data and model parallelism. We pre-trained for 2 epochs and fine-tuned for 10 epochs.

### Hyperparameters

**Continous pretraining on Stack-SQL dataset**

- Hardware: SambaNova Reconfigurable Dataflow Unit (RDU)
- Optimizer: AdamW
- Epochs: 2
- Global Batch size: 256
- Batch tokens: 256 * 4096 = 1,048,576 tokens
- Learning Rate: 1e-5
- Learning Rate Scheduler: Fixed
- Warmup Steps: 0
- Weight decay: 0.1

**Finetuning on NSText2SQL dataset**

- Hardware: SambaNova Reconfigurable Dataflow Unit (RDU)
- Optimizer: AdamW
- Epochs: 10
- Global Batch size: 64
- Batch tokens: 64 * 4096 = 262,144 tokens
- Learning Rate: 1e-5
- Learning Rate Scheduler: Cosine Schedule with Warmup
- Warmup Steps: 0
- End Learning Ratio: 0.1
- Weight decay: 0.1
## Intended Use and Limitations

The model was designed for text-to-SQL generation tasks from given table schema and natural language prompts. The model works best with the prompt format defined below and outputting `SELECT` queries.

## How to Use

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("sambanovasystems/SambaCoder-nsql-llama-2-70b")
model = AutoModelForCausalLM.from_pretrained("sambanovasystems/SambaCoder-nsql-llama-2-70b", torch_dtype=torch.bfloat16)
text = "CREATE TABLE stadium (
    stadium_id number,
    location text,
    name text,
    capacity number,
    highest number,
    lowest number,
    average number
)

CREATE TABLE singer (
    singer_id number,
    name text,
    country text,
    song_name text,
    song_release_year text,
    age number,
    is_male others
)

CREATE TABLE concert (
    concert_id number,
    concert_name text,
    theme text,
    stadium_id text,
    year text
)

CREATE TABLE singer_in_concert (
    concert_id number,
    singer_id text
)


-- Using valid SQLite, answer the following questions for the tables provided above.

-- What is the average, minimum, and maximum age of all singers from France?
SELECT"
input_ids = tokenizer(text, return_tensors="pt").input_ids

generated_ids = model.generate(input_ids, max_length=500)
print(tokenizer.decode(generated_ids[0], skip_special_tokens=True))
```