Tune the Rag

Convert any Huggingface dataset to a retrieval-augmented dataset
Read the blog post »

About The Project

• Use this repo to augment prompts with semantically similar prompts from the training set.
• Choose the embedding model to perform the semantic search.
• Customise the prompts.

For example, the data point from an SQL dataset {"question": "How many singers do we have?", "context": "CREATE TABLE singer (Id VARCHAR)"} gets prompted to be:

You are a powerful text-to-SQL model. Your job is to answer questions about a database. You are given a question and context regarding one or more tables. You must output the SQL query that answers the question.

Given the following example:
### Input:
How many artists do we have?

### Context:
CREATE TABLE artist (Id VARCHAR)

### Response:
SELECT count(*) FROM artist

Please generate the SQL query that answers the following:
### Input:
How many singers do we have?

### Context:
CREATE TABLE singer (Id VARCHAR)

### Response:

And these are the results of using retrieval-augmented prompts vs few-shot prompts:

Getting Started

Prerequisites

1. Clone repo

git clone https://github.com/samlhuillier/tunetherag.git

2. Install requirements

pip install -r requirements.txt

Creating a RAG dataset

1. Open tunetherag.ipynb
2. In the third cell, modify the dataset loading arguments:

   embedding_feature = "question"
   dataset_parameters = {"dataset_name": "gsm8k", "config_name": "main"}

   (embedding_feature is what will be embedded and dataset_parameters will be plugged into load_dataset)
3. Setup prompts:

   def format_math_example(example):
    inference_prompt = f"""### Problem:
    {example["question"]}
   
    ### Answer:"""
   
    full_prompt = f"{inference_prompt}\n{example['answer']}"
   
    return full_prompt, inference_prompt
   
   math_prompt = "Solve the following math problem thinking step-by-step:"

4. Run cells to generate the retrieval augmented dataset which will be saved as a set of json files per split!

(back to top)

License

Distributed under the MIT License.