search

allenai / open-instruct

Apache License 2.0
1.08k stars 140 forks source link

Wandb updates: Separate DPO logging from IFT + add team option #160

Closed natolambert closed 1 month ago

natolambert commented 2 months ago

Will test this shortly.

hamishivi commented 1 month ago

Did you end up testing this?

natolambert commented 1 month ago

@hamishivi yup, just launched a bogus job to immediately error post wandb init, but here's my new project https://wandb.ai/natolambert/open_instruct_dpo

natolambert commented 1 month ago

Also, with --wandb_entity rlhf-llm-dev https://wandb.ai/rlhf-llm-dev/open_instruct_dpo

hamishivi commented 1 month ago

LGTM! Seems like your run hit some error we might need to work out at some point, but unrelated to this change.

natolambert commented 1 month ago

Yeah @hamishivi there was a llama 2 dpo tokenizer error that I was confused by, the assert was failing, but I commented it out to just get things running quickly to make sure this works.

Will test more soon.