---
license: llama2
inference:
  parameters:
    do_sample: false
    max_length: 200
widget:
- text: "CREATE TABLE stadium (\n    stadium_id number,\n    location text,\n    name text,\n    capacity number,\n)\n\n-- Using valid SQLite, answer the following questions for the tables provided above.\n\n-- how many stadiums in total?\n\nSELECT"
  example_title: "Number stadiums"
- text: "CREATE TABLE work_orders ( ID NUMBER, CREATED_AT TEXT, COST FLOAT, INVOICE_AMOUNT FLOAT, IS_DUE BOOLEAN, IS_OPEN BOOLEAN, IS_OVERDUE BOOLEAN, COUNTRY_NAME TEXT, )\n\n-- Using valid SQLite, answer the following questions for the tables provided above.\n\n-- how many work orders are open?\n\nSELECT"
  example_title: "Open work orders"
- text: "CREATE TABLE stadium ( stadium_id number, location text, name text, capacity number, highest number, lowest number, average number )\n\nCREATE TABLE singer ( singer_id number, name text, country text, song_name text, song_release_year text, age number, is_male others )\n\nCREATE TABLE concert ( concert_id number, concert_name text, theme text, stadium_id text, year text )\n\nCREATE TABLE singer_in_concert ( concert_id number, singer_id text )\n\n-- Using valid SQLite, answer the following questions for the tables provided above.\n\n-- What is the maximum, the average, and the minimum capacity of stadiums ?\n\nSELECT"
  example_title: "Stadium capacity"
---

# 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, 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.

### Basic Information

<!-- Provide the basic links for the model. -->
- **Blog Post**: [Link](TBA)
- **Discord**: [Link](TBA)
- **HF Hosting**: [Chat with me!](TBA)

## 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

NSQL 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.

## 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

Example 1:

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("sambanovasystems/nsql-Llama-2-70B")
model = AutoModelForCausalLM.from_pretrained("sambanovasystems/nsql-Llama-2-70B", torch_dtype=torch.bfloat16)

```