Popular Open-Source Large Language Models

Here are the popular and powerful open-source large language models to consider in 2023

Open-source large language models, like GPT-3.5, are cutting-edge AI programs created to comprehend and produce text that resembles that of a person based on the patterns and knowledge they have gained from massive training data. These models are created using deep learning techniques, and their training data consists of enormous datasets with a wide variety of text sources, such as books, articles, webpages, and other textual materials.

The phrase "open source" refers to the model's code and underlying architecture being made accessible to the general public, enabling developers and researchers to use, improve, and alter the model for various uses. This transparency encourages cooperation and innovation within the AI community, helping people and organizations to improve on pre-existing models, develop fresh applications, and advance AI technology.

Numerous interconnected neural network layers in large language models like GPT-3.5 process and analyze text data. The models develop their ability to recognize patterns, comprehend syntax and semantics, and produce coherent and contextually appropriate responses based on input throughout training.

GPT-3 & GPT-4 by OpenAI

The highly large language model GPT-3/4 (Generative Pre-trained Transformer 3/4) was created by OpenAI. The GPT series' third iteration has received great praise and attention in artificial intelligence (AI) and natural language processing (NLP).

LaMDA by Google

Google created the conversational Large Language Model (LLM) known as LaMDA AI, which stands for Language Model for Dialogue Application, as the core technology for apps that use dialogue and can produce human-sounding language. LaMDA, a work in natural language processing that forms the foundation for various language models, notably GPT-3, the technology powering ChatGPT, is one of the innovations from Google's Transformer research project.

LLaMA by Meta AI

Large Language Model Meta AI, or LLaMA for short, is a large language model (LLM) that Meta AI announced in February 2023. Models with sizes ranging from 7 billion to 65 billion parameters were trained. The performance of the 13 billion parameters LLaMA model on most NLP benchmarks outperformed that of the much bigger GPT-3 (with 175 billion parameters), according to the model's creators.

Bloom by BigScience

The Big Open-science Open-access Multilingual Language Model (BLOOM) is a significant language model BigScience developed based on transformers. More than 1000 AI researchers created it to provide open access to a substantial language model for anyone who wants to use it.

PaLM by Google

Palm, a large language model with 540 billion parameter transformers, was developed by Google AI. Researchers also trained PaLM models with 8 and 62 billion parameters to assess the effects of the model scale. Translation, code production, humor explanation, common sense, and mathematical reasoning are just a few of the tasks PaLM is capable of.

Dolly by Databricks

The Databricks machine-learning platform was used to train Dolly, a large language model that learns to obey commands. It was trained using roughly 15k instruction/response fine-tuning records based on Pythia-12b, including brainstorming, categorization, closed QA, generation, information extraction, open QA, and summarization produced by Databricks personnel.

Cerebras-GPT from Cerebras

To promote research into LLM scaling laws using open architectures and data sets and to show how easy and scalable it is to train LLMs on the Cerebras hardware and software stack, the Cerebras-GPT family has been made available. Hugging Face has all Cerebras-GPT variants available.

BERT by Google

Researchers at Google AI developed the well-known language model BERT (Bidirectional Encoder Representations from Transformers) in 2018. It has considerably impacted several downstream jobs and the natural language processing (NLP) field.

XLNet by Google

A language model called XLNet was released in 2019 by Google AI researchers. It overcomes the drawbacks of conventional language models, such as left-to-right or auto-regressive pre-training methods.