Mr_TJ

icon location An explorer. Email :[email protected]

Results 5 issues of Mr_TJ

feature(tj): add evaluation utils from open-reasoner-zero to priorzero

./zoo/jericho/priorzero/orz-eval ./zoo/jericho/priorzero/priorzero_entry.py

research

feature(tj): add monitoring for the gradient conflict metric of MoE in ScaleZero

feature(tj): add monitoring for the gradient conflict metric of MoE in ScaleZero

config
research

fix(tj): add moe grad analysis toy example

根据之前pr的要求 进行修改

config

fix(tj): finetune spaceinvaders from atari26 pretrained ckpt in ScaleZero

atari_unizero_multitask_segment_ddp_config_debug_naive.py 单任务从0微调 atari_unizero_multitask_segment_ddp_config_finetune_SpaceInvaders_full.py 全量微调 atari_unizero_multitask_segment_ddp_config_finetune_SpaceInvaders_head_back_encoder_lora.py 微调head+encoder(lora)+backbone(lora) atari_unizero_multitask_segment_ddp_config_finetune_SpaceInvaders_head_back_lora.py 微调head+backbone(lora) atari_unizero_multitask_segment_ddp_config_finetune_SpaceInvaders_head.py 微调(head)

enhancement
research