Post training examples - Githubissues

EleutherAI / gpt-neox

An implementation of model parallel autoregressive transformers on GPUs, based on the Megatron and DeepSpeed libraries

https://www.eleuther.ai/

Apache License 2.0

6.95k stars 1.02k forks source link

Closed dmahan93 closed 2 months ago

dmahan93 commented 2 months ago

recreate zephyr 7b beta from the sft https://huggingface.co/HuggingFaceH4/mistral-7b-sft-beta
llama 3 configs for sft/dpo/rm on ultrafeedback
implementation improvements that came up when poking at it (cross entropy is negative logprobs, and we have an efficient cross entropy! no need to do a full gather over the logits 🚀)

dmahan93 commented 2 months ago

will need to edit the configs once whatever changes in https://github.com/EleutherAI/gpt-neox/pull/1276 get finalized

Quentin-Anthony commented 2 months ago

will need to edit the configs once whatever changes in #1276 get finalized

It's been merged. You're clear to do a final pass + test run.

Quentin-Anthony commented 2 months ago

Tested and working for me