open-llm-search

Model Overview

As the demand for large language models grows, a common limitation surfaces: their inability to directly search the internet. Although tech giants like Google (with Bard), Bing, and Perplexity are addressing this challenge, their proprietary methods have data logging issues.

Introducing Open LLM Search — A specialized adaptation of Together AI's llama-2-7b-32k model, purpose-built for extracting information from web pages. While the model only has a 7 billion parameters, its fine-tuned capabilities and expanded context limit enable it to excel in search tasks.

License: This model uses Meta's Llama 2 license.

Fine-Tuning Process

The model's fine tuning involved a combination of GPT-4 and GPT-4-32k to generate synthetic data. Here is the training workflow used:

1. Use GPT-4 to generate a multitude of queries.
2. For each query, identify the top five website results from Google.
3. Extract content from these websites and use GPT-4-32k for their summarization.
4. Record the text and summarizes from GPT-4-32k for fine-tuning.
5. Feed the summaries from all five sources with GPT-4 to craft a cohesive response.
6. Document both the input and output from GPT-4 for fine-tuning.

Fine tuning was done with an <instructions>:, <user>:, and <assistant>: format.

Getting Started

• Experience it firsthand! Check out the live demo here.
• For DIY enthusiasts, explore or self-deploy this solution using our GitHub repository.