HACKER Q&A
📣 baobabKoodaa

Open-source ChatGPT alternatives?


What's the state of the art in open source GPT models right now, in practical terms? If your typical use case is taking a pretrained model and fine tuning it to a specific task, which LLM would yield the best results while running on consumer hardware? Note that I'm specifically asking for software that I can run on my own hardware, I'm not interested in paying OpenAI $0.02 per API request.

I'll start the recommendations with Karpathy's nanoGPT: https://github.com/karpathy/nanoGPT

What else do we have?


  👤 valgaze Accepted Answer ✓
Clue on conversation "history"-- "While ChatGPT is able to remember what the user has said earlier in the conversation, there is a limit to how much information it can retain. The model is able to reference up to approximately 3000 words (or 4000 tokens) from the current conversation - any information beyond that is not stored.

Please note that ChatGPT is not able to access past conversations to inform its responses."

https://help.openai.com/en/articles/6787051-does-chatgpt-rem...

Some interesting techniques I've seen involve essentially a ring-buffer and after each turn a call is made to summarize the conversation up to that point and use that as context for subsequent prompt


👤 smoldesu
This runs fine in RAM constrained (<2gb) situations: https://huggingface.co/EleutherAI/gpt-neo-125M

It's bigger brother, 1.3b, uses ~5.5gb of memory but yields slightly more GPT-like answers. Both take ~5-20 seconds to generate a response though, so take that into account when building with it.


👤 speedgoose
You also have GPT J 6B and BLOOM but to be honest they are not like ChatGPT.

https://huggingface.co/EleutherAI/gpt-j-6B

https://huggingface.co/bigscience/bloom


👤 SparkyMcUnicorn
Open Assistant (started by some of the people that started Stable Diffusion I think?) is very early, but looks very promising.

https://open-assistant.io/

https://github.com/LAION-AI/Open-Assistant


👤 gigel82
GPT Neo 1.3B (https://huggingface.co/EleutherAI/gpt-neo-1.3B) is the largest I can run on my 12Gb VRAM GPU, and I'm sorry to say it's output is a joke (nowhere near GPT-3, more like GPT-2 level of BS).

However, you can fine tune it; and I'm sure with lots of fine tuning and some jiggling of the parameters you can get a half decent custom-purpose solution.


👤 lopuhin
In terms of models which are reasonably fast to run and easy to install, I think Flan-T5 is one of the best: https://huggingface.co/google/flan-t5-xxl - although out of the box it's more focused on giving short answers and it's very far from ChatGPT.


👤 v3ss0n
Tried all of them, they are still at gpt2 tier. None of them close to gpt3, so far away from davinci.

To run your own laptop and cheap we still need much fine tuned training set with much better algorithm .

Right now, most capable ones needs over 120gb of VRAM just inference (run).


👤 rvz
We seriously need more mature and open-source LLMs that are as good as, if not better than ChatGPT. It was a great surprise to see Stable Diffusion disrupt DALLE-2 with open source

We need the same for ChatGPT and GPT-4.


👤 moomoo11
I’m only a consumer of GPT.

Any pointers as to where someone like me with software engineering experience but literally no AI knowledge can train my own GPT on my own data sets?

In my case I have downloaded some public domain databases (1-7gb each) and I would like to get some additional insights out of them. I have been querying them and using them to build my company but I’m curious to know if GPT can help me in that regard.


👤 juliensalinas
Some alternatives are explained in this article: https://nlpcloud.com/chatgpt-open-source-alternatives.html But it will require some prompt engineering in order to get the same level of instruction as ChatGPT.

👤 PaulHoule
Here is a Python package that can download transformer embeddings automatically

https://www.trychroma.com/

In general a lot of people download models from huggingface, I think that package automates that task.


👤 vvipgupta
You might also want to check out https://github.com/lucidrains/PaLM-rlhf-pytorch

👤 dieselgate
I'm not very familiar with this space but would have thought "OpenAI" would be at least somewhat open-source. Is this just naming and not relevant to the product at all?