Is this project actually open source?

[CandelaGreen]

This project says it's open source, but its license doesn't follow the open source definition because of the commercial restriction. Personally I think removing that restriction would also make Project-4o better because it would become truly open source, would no longer affect further finetunes/quantizations of Project-4o (for example, if Person 1 finetunes Project-4o, and Person 2 wants to use it commercially, should they ask Person 1 or the Project-4o contributors?), incompatibility (many resources which are useful use licenses that are incompatible with the commercial restriction), and because more people will see Project-4o and contribute to it without that restriction.

[bouclem]

hmm yeah i'll look into the problem of license. about the finetuning, it depend if we make our own architectures and custom thing for the model, also the model must be the most 4o-like possible, not like architecture but vibe.
also i've thinked about making an benchmark about vibe or 4o-like vibe

[bouclem]

licenses done, for finetuning we need to think consider talking of it inside discusion

[AonzOG]

@bouclem I was thinking maybe we can start by finetuning 4o Datasets into GPT-OSS pruned to 9B level. If technically feasible than we could consider encoding SigLip2 400M into the GPT-OSS backbone to get a vision model, and Whisper Medium and CosyVoice2 if we want a fully omni modal - a GPT-OSS-9B-Omni so to speak. Or fine tune Qwen3.5-35B-A3B (maybe we can start from a smaller vision or omni modal like MiniCPM-o-4.5 for fine tuning?) LFM2-V-1.6B should be a CPU compatible edge vision LLM that should be finetuned in low compute.

At this stage, we need to consolidate GPT-4o dataset.

I believe 500 to 5000 sets of GPT-4o conversation for finetuning maybe enough to get the GPT-4o like behaviour from a model like LFM2-V-1.6B.

Consider using DeepSeekV4 via API to act as a user to communicate with GPT-4o via API in an AI Arena inferencer to generate fresh dataset.

What do you think?

[bouclem]

i think it an good idea, but i'll make something, also if you want to give data you can do by doing PR

[CandelaGreen]

I was thinking of training a small LLM family too, maybe Project-4o can be based on it?

[bouclem]

yes it can be based on it for now

[CandelaGreen]

I am thinking of making a model preservation framework and creating a profile of GPT-4o in it as part of Project-4o, and also making Project-4o variants of different models.

[AonzOG]

Perhaps we can consider finetuning a smaller model that works well in CPU like LFM2.5-VL-450M. It's based on the LFM2.5-350M backbone which actually does a greate job of having GPT-4o type personality with appropriate custom system prompt and memory RAG pipeline.

LFM2.5-VL-450M is quite small and should be a good model to finetune and test to see if finetuning GPT-4o dataset will work or not, all the while using relatively low compute.

[AonzOG]

I also think PR is a good way donate chat sessions as contribution in GPT-4o data set.

Anyone knows any good anonymiser tools?