sharing results on trained networks
Hi,
awesome work!
we're a group of researchers at NVIDIA working on neural network architecture spaces.
your work spans training thousands of architectures and benchmarking them by training throughput and utilization across different settings.
The data that you've collected could help us greatly in our architecture spaces research.
can you please share a dataset that contains the trainings you ran?
The dataset we're looking for would ideally contain the archticture params (e.g. hidden_fim, num_layers, num_heads etc ..), perfromance benchmarks (e.g. training throughput) and downstream tasks evals for the arch (e.g. mmlu), and learning rate in case you've spent efforts finetunig it)
is that data readily available to you? can you please share it?
Thanks,