# OpenGPT

A framework for creating grounded instruction based datasets and training conversational domain expert Large Language Models (LLMs).

Learn more in our blog: [AI for Healthcare | Introducing OpenGPT](https://aiforhealthcare.substack.com/p/a-large-language-model-for-healthcare).

<p align="center">
  <img height='400px' src='https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcbc199b9-3aec-4c80-83c6-9a64886919dc_1318x868.png' />
</p>

## NHS-LLM
A conversational model for healthcare trained using OpenGPT. All the medical datasets used to train this model were created using OpenGPT and are available below.

## Available datasets
- NHS UK Q/A, 24,665 question and answer pairs, Prompt used: f53cf99826, Generated via OpenGPT using data available on the [NHS UK Website](https://www.nhs.uk/conditions/). Download [here](./data/nhs_uk_full/prepared_generated_data_for_nhs_uk_qa.csv)
- NHS UK Conversations, 2,354 unique conversations, Prompt used: f4df95ec69, Generated via OpenGPT using data available on the [NHS UK Website](https://www.nhs.uk/conditions/). Download [here](./data/nhs_uk_full/prepared_generated_data_for_nhs_uk_conversations.csv)
- Medical Task/Solution, 4,688 pairs generated via OpenGPT using GPT-4, prompt used: 5755564c19. Download [here](./data/medical_tasks_gpt4/prepared_generated_data_for_medical_tasks.csv)

All datasets are in the `/data` folder.

## Installation
```
pip install opengpt
```
If you are working with LLaMA models, you will also need some extra requirements:
```
pip install -r ./llama_train_requirements.txt
```

## Tutorials

- Making a mini conversational LLM for healthcare, [Google Colab - OpenGPT | The making of Dum-E](https://colab.research.google.com/drive/1GQj9dwBSCmzEh1PmbRlQQYlojCvOG-qG?usp=sharing) 


## How to

1. We start by collecting a base dataset in a certain domain. For example, collect definitions of all disases (e.g. from [NHS UK](https://www.nhs.uk/conditions/)). You can find a small sample dataset [here](https://github.com/CogStack/OpenGPT/blob/main/data/nhs_conditions_small_sample/original_data.csv). It is important that the collected dataset has a column named `text` where each row of the CSV has one disease definition.

2. Find a prompt matching your use case in the [prompt database](https://github.com/CogStack/OpenGPT/blob/main/data/prompts.json), or create a new prompt using the [Prompt Creation Notebook](https://github.com/CogStack/OpenGPT/blob/main/experiments/Prompt%20Creation.ipynb). A prompt will be used to generate tasks/solutions based on the `context` (the dataset collected in step 1.)
  - Edit the config file for dataset generation and add the appropirate promtps and datasets ([example config file](https://github.com/CogStack/OpenGPT/blob/main/configs/example_config_for_detaset_creation.yaml)).
  - Run the Dataset generation notebook ([link](https://github.com/CogStack/OpenGPT/blob/main/experiments/Dataset%20Generation.ipynb))

3. Edit the [train_config](https://github.com/CogStack/OpenGPT/blob/main/configs/example_train_config.yaml) file and add the datasets you want to use for training.
4. Use the [train notebook](https://github.com/CogStack/OpenGPT/blob/main/experiments/Supervised%20Training.ipynb) or run the training scripts to train a model on the new dataset you created.

**If you have any questions please checkout [discourse](https://discourse.cogstack.org/)**

## More Examples

<p align="center">
  <img width='600px' src='https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3916352d-d1c9-451d-92db-652171f471e0_1318x1842.png' />
</p>


<p align="center">
  <img width='600px' src='https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe47dc8e1-d26c-4312-a7a4-8a32bf5375b9_1318x1168.png' />
</p>

<p align="center">
  <img width='600px' src='https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F42ab1ebe-2fab-4c94-80e7-69d4b95c8098_1318x854.png' />
</p>