Conversational AI, such as virtual agents and chatbots, use natural language processing and vast amounts of data to imitate human interactions and understand speech and text. The field of conversational AI has seen significant advancements in recent years, with the introduction of ChatGPT. Let’s explore some other game-changing open-source large language models (LLMs) in conversational AI.
LLaMa: Developed by Meta AI, LLaMa is a versatile and responsible LLM. Its release aims to make AI research more accessible and promote responsible AI practices. LLaMa comes in different sizes, with parameters ranging from 7B to 65B. Access to the model is granted on a case-by-case basis to industry research labs, academic researchers, and others.
Open Assistant: LAION-AI developed Open Assistant as a chat-based LLM that can perform various tasks, including answering queries, generating text, translating languages, and creating content. Despite being in the development stage, Open Assistant already possesses skills like interacting with external systems, such as Google Search. It is an open-source initiative, allowing contributions from anyone.
Dolly: Databricks created Dolly, an instruction-following LLM powered by the Pythia 12B model. Trained on approximately 15k instruction/response records, Dolly excels in following instructions with impressive accuracy, despite not being cutting-edge.
Alpaca: Stanford University developed Alpaca as a cost-effective instruction-following model based on Meta’s LLaMa (7B parameters) model. It performs well on various instruction-following tasks and costs less than $600 to produce. Trained on 52,000 demonstrations, Alpaca resembles OpenAI’s text-davinci-003 model but is more affordable.
Vicuna: Vicuna is a chatbot developed by a collaborative team from UC Berkeley, CMU, Stanford, and UC San Diego. Fine-tuned on LLaMa using conversations from ShareGPT, Vicuna offers engaging and natural conversation capabilities. With 13B parameters, it provides detailed and well-structured answers comparable to ChatGPT.
Koala: Developed by the Berkeley Artificial Intelligence Research Lab (BAIR), Koala is a dialogue model based on the LLaMa 13B model. It prioritizes safety and interpretability, making it suitable for studying language model safety and bias. Koala is an open-source alternative to ChatGPT and includes EasyLM, a framework for training and fine-tuning LLMs.
Pythia: Eleuther AI introduces Pythia, a set of autoregressive language models specifically designed for scientific research. With models ranging from 70M to 12B parameters, Pythia enables comparisons and exploration of scaling effects.
OpenChatKit: Developed by Together, OpenChatKit is an open-source chatbot development framework that simplifies the process of building conversational AI applications. Built on the GPT-4 architecture, OpenChatKit is available in three model sizes to accommodate different computational resources and application requirements.
RedPajama: RedPajama is a collaborative project aiming to develop cutting-edge open-source models. Starting with reproducing the LLaMa training dataset, the project aims to create an open and replicable language model. The dataset is accessible through Hugging Face, and users can replicate the results using Apache 2.0 scripts available on GitHub.
StableLM: Stability AI presents StableLM, an open-source language model trained on a dataset three times larger than The Pile dataset. Despite its small size, StableLM performs well in conversational and coding tasks. It comes in 3B and 7B parameters and can generate text and code for various applications.
These models revolutionize the field of conversational AI and provide researchers and developers with powerful tools to create innovative applications. Collaboration and open-source initiatives contribute to their continuous improvement. For more information on each model, refer to the provided references. Join the ML SubReddit, Discord Channel, and Email Newsletter for the latest AI research news and exciting projects.