Report Different models fight on the street

[taozhiyuai]

two models fight 50 rounds, the report is below

[Nicolas Oulianov]

Impressive results ! Finally a benchmark where Gemma wins lol

You should have a file called "results.csv", right ? Is it the one you used to compute the win rates ?

[taozhiyuai]

yes, the data in the table are all from results.cvs

[taozhiyuai]

Impressive results ! Finally a benchmark where Gemma wins lol

You should have a file called "results.csv", right ? Is it the one you used to compute the win rates ?

I try to choose the same size of model parameters, or the same file size of model with same Q level. try to keep similar speed of token generation. big model always fail because of low speed of token generation.

I think Gemma 7b is good enough , it is time to train the model.

[Nicolas Oulianov]

You want to do finetuning ?

[taozhiyuai]

You want to do finetuning ?

yes, it is interesting.