was there a purpose to loading the model in 8 bits, comparing it to 32b llama and then saying the comparison was inaccurate?
#3
by Alignment-Lab-AI - opened
Sorry, did we write anything about the inaccuracy of the comparison anywhere?
Yes
If you are talking about this:
Then, we miss the comparison with base Llama models loaded in 8bits. I didn't take the time to do it when publishing the models.
If you have insights about that benchmark, open a PR and add them to the benchmark table.
Otherwise, did you try the model?
chainyo changed discussion status to closed