recursive json convert chatgpt

robust get_expert_stats and remove repeat

everything into the expertracking class

attribute error

handle JSON serialization

claude to deepseek

Revert "expert usage stats"

HFLM forward to loglikelihoods

add evalexperts

fix hflm pretrained in eval

Revert "18k checkpoint"

attempts to fix more

run_evaluation fix

load custom model fix

more debug info

load custom model fix

all OlMoeConfig to MyOlmoeConfig

myolmoe name again

use olmoeconfig

olmoe to myolmoe

claude's eval overhaul

add eval debug

use path for HFLM

Checkpoint at step 12000

keeping save steps at 2k

save steps to 5

changes to training script

progress 20k

Checkpoint at step 2000

delete checkpoints all epochs to 3

trying changes to git stuff

save functionality

YOLO run setup

attempt new distribution experts

batch 2 is biggest

syntax

batch 3

batch size 2

batch 4

comma

16batch

train agaaa

train aga

fix train

debugging missing grad

update training script

reorder small experts

unfreeze only gate and experts

1 batch size

batch size back to 2