ChatGPT is built on LLM and Framework as we all know. In today's topic, we will introduce eight open source frameworks and models that you can use to build your own chatbot or AI product.

1. LLaMA

The LLaMA project is based on the Fundamental Language Model. which has 7 billion to 65 billion parameters. These models are trained on millions of tokens and publicly available datasets. As a result, LLaMA-13B outperforms GPT-3 (175B) and can show similar performance to models like LLaMA-65B, Chinchilla-70B and PaLM-540B.

GitHub: facebookresearch/llama

Demo:  Baize Lora 7B

2. Alpaca

Stanford Alpaca claims they can compete with ChatGPT and anyone can build their own ChatGPT for less than $600. The Alpaca 7B is fine-tuned from the LLaMA 7B model based on 52,000 instructions.

GitHub: tatsu-lab/stanford_alpaca

Demo: Alpaca-LoRA

3. Vicuna

Vicuna is fine-tuned from the LLaMA model, based on user shared conversations collected from ShareGPT. Vicuna-13B has already achieved 90% quality efficiency from OpenAI's ChatGPT and Google Bard. It outperforms the LLaMA and Stanford Alpaca models in about 90% of cases. Vicuna cost about $300 to train.

GitHub: lm-sys/FastChat

Demo: FastChat (lmsys.org)

4. OpenChatKit

OpenChatKit is an open source ChatGPT alternative complete toolkit with which you can build your own chatbot. It will provide instructions for training and fine-tuning your model. OpenChatKit will help you create chatbots after all. The GPT-NeoXT-Chat-Base-20B model already outperforms GPT-NoeX in query-response.

GitHub: togethercomputer/OpenChatKit

Demo: OpenChatKit 

Model card: togethercomputer/GPT-NeoXT-Chat-Base-20B

5. GPT4ALL

GPT4ALL is a community driven project trained on massive data such as code, assistant interactions, stories, etc. The team behind this model has made their dataset, model volume, data curation process, training code all open source. They also released a 4 bit version of this model that you can run on a laptop. You can also run this model in a Python client if you want.

GitHub: nomic-ai/gpt4al

Demo: GPT4All

Model card: nomic-ai/gpt4all-lora · Hugging Face

6. Raven RWKV

Raven RWKV 7B is an open source chatbot built on the RWKV language model that can produce results similar to ChatGPT. This model uses RNNs that match Transformers in terms of quality and scaling, and saves a lot of VRAM. Raven, fine-tuned from Stanford Alpaca, code-alpaca, and many other datasets.

GitHub: BlinkDL/ChatRWKV

Demo:  Raven RWKV 7B

Model card: BlinkDL/rwkv-4-raven

7. OPT

OPT or Open Pre-trained Transformer language model is not better than ChatGPT but it has shown great performance in Stereotypical Bias Analysis. You can integrate Alpa, Colossal-AI, CTranslate2, and FasterTransformer with it to get even better results.

It is on this list because its monthly downloads are like 624710 in text generation category.

GitHub: facebookresearch/metaseq

Demo: A Watermark for LLMs

Model card: facebook/opt-1.3b

8. Flan-T5-XXL

The Flan-T5-XXL is a finetune T5 model. Here is a kind of instruction based fine tuning. Such instructional fine tuning has dramatically increased the performance of models such as PaLM, T5, and U-PaLM. The Flan-T5-XXL model has also been fine-tuned with over 1,000 additional tasks that further enhance its performance.

GitHub: google-research/t5x

Demo: Chat Llm Streaming

Model card: google/flan-t5-xxl

last word

There are so many open source options available on the internet that I tried to highlight some of the best. These open source chatbots or models will get better in the next few months then it might even overtake ChatGPT in terms of performance.

So today until the next topic stay well.