---
license: mit
language:
- en
metrics:
- accuracy
pipeline_tag: text-generation
widget:
- text: "<schema>CREATE TABLE radio(age VARCHAR, radio_id VARCHAR,  frequency VARCHAR, wavelength VARCHAR); CREATE TABLE  radio_faults(radio_id VARCHAR, fault_description VARCHAR)</schema><question>Get the radio id and defect descriptions of radios that have wavelength greater than 30 ?</question><sql>"
  example_title: "example1"
- text: "<schema>CREATE TABLE system(JobID: String,GID: String, UID: String, Start:Time(yyyy/mm/dd), End: Time,ElapsedRaw: Time, CPUTimeRAW: Time,NCPUS: Number,NNodes: Number, NodeList: List,  State:String, Timelimit: Time);</schema><question>Get UID and job id for Jobs that started on Jan 20 , 2023</question><sql>"
  example_title: "example2"
- text: "<schema>CREATE TABLE department (Department_ID number, Name text, Creation text, Ranking number, Budget_in_Billions number, Num_Employees number) which has Department_ID as primary key abd CREATE TABLE head (head_ID number, name text, born_state text, age number) which has head_ID as primary key and CREATE TABLE management (department_ID number, head_ID number, temporary_acting text) which has department_ID as primary key</schema><question>"
  example_title: "example3"
tags:
- code
- sql
- text2sql
- instruction_tuned
- jax
- pytorch
- 1b
- expert
datasets:
- PipableAI/spider-bird
---
# Pipable’s pipSQL
Please refer to https://huggingface.co/PipableAI/pipSQL-1.3b for our state of the art model, that gives better performance than chatgpt and claude on sql tasks on a lot of benchmarks.


Pipable’s pipSQL is a model distilled from llama 1b to generate sql queries given prompt and schema. 
We used a unique pipeline which involved the model working on two objectives alternatively ----
1. Maximizing the log prob of all tokens in the sequence (including the prompt tokens) 
2. Minimizng the difference between the true value and the predicted maximum value of the output tokens i.e generated tokens for the sql query slice of the entire sequence.

During our research on training models of different param sizes, we have seen recurring pattern where zero shot inference capabilities on questions that need decuctive reasoning, is an attribute that models with params higher than 2.7B exhibit.
pip-SQL-1B is demonstrative, refer to pip-SQL-7B and pip-SQL-3B(coming soon)for SOTA models open sourced by pipable



## License

The model's new weights along with all other assets involved with it are open sourced under mit license.

## How to Use

```python
text = """<schema>{schema}</schema>
<question>{question}</question>
<sql>"""
```
pytorch

```python
from transformers import AutoModelForCasualLM, AutoTokenizer
device = "cuda"
model = AutoModelForCausalLM.from_pretrained("PipableAI/pipSQL1b")
tokenizer = AutoTokenizer.from_pretrained("PipableAI/pipSQL1b")

inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True).split('<sql>')[1].split('</sql>')[0])
```
flax

```python
from transformers import FlaxAutoModelForCasualLM, AutoTokenizer
model = FlaxAutoModelForCausalLM.from_pretrained("PipableAI/pipSQL1b" , from_pt=True)
tokenizer = AutoTokenizer.from_pretrained("PipableAI/pipSQL1b")
```

## The PipableAI team

Avi Kothari, Pratham Gupta, Ritvik Aryan Kalra, Rohan Bhatial, Soham Acharya