File size: 3,880 Bytes
4acbe3a b6a8eb3 216215c b6a8eb3 216215c c0f3ff6 216215c 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](TBA)
- **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))
``` |