Cong Lu
Cong Lu
We would advise sticking to GPT-4o but trying to increase your OpenAI account tier.
Hey! This code was only tested using NVIDIA GPUs, I believe they have a guide for PyTorch here: https://developer.apple.com/metal/pytorch/ :)
These prepare the data for the nanogpt runs. The remaining commands run the baseline which should be machine dependent for things like training speed!
I think modern machine learning is quite hard without a GPU. Later parts of the pipeline will attempt dozens of runs which could take hours each without a GPU. I...
> Why not use GPT4o-mini/Claude instead of a local nanogpt? Not totally sure of the value here for a hybrid approach given the cost of cutting off mac users since...
> Additionally, is it REALLY a dependency? It looks like it creates an artifact that is used later when you’re using the GPT4o-mini. > > Is that artifact actually a...
Hey! We have automatic exponential backoff on this but we do recommend applying for higher rate limit API access if parallelising heavily :)
Q1. Likely the most effective route would be to provide some of the base code for the above in the template. Q2. It can fetch datasets it knows the names...
This should be solved by: https://github.com/SakanaAI/AI-Scientist/issues/89#issuecomment-2323129188
Hey! This should be extremely easy to add, essentially you just need to modify the LLMs list and route the requests appropriately in: https://github.com/conglu1997/ACD/blob/main/src/llm_utils.py