Repository Issue Activity (beta)

#365 Time to switch off 8xH100?

#355 Track 3: two small logging gaps that make a record hard to verify from its logfile

#174 Dataloader of the training data doesn't shuffle and Fineweb subsets aren't pre-shuffled. Use fineweb_100BT-shuffled / fineweb_edu_100BT-shuffled if you need to

#258 We should use official FA3 builds now

#296 Track 3: test Kaon as an optimizer-only Muon ablation

#263 Why not use DDP on Track_3_Optimization ?

#268 Nan error during multiple runs.

#223 Question: are we not cheating by performing (by now thousands or tens of thousands) of architecture and hyperparameter iterations over the same val set?

#256 parallel subspace unembed and block carvouts

#245 Feature request : support for vision models in the future

#234 Use just 1 lambda for input on layer 0

#142 Add nanochat track

#220 A new optimizer that is (allegedly) faster than Muon

#115 Code doesn't actually work for fewer GPUs

#182 Time to update the description!

#210 Has anyone tried the LFM2 architecture?

#184 Train_switch

#203 will someone test this?

#198 Speed run is currently broken?

#173 Why didn't we use 1e-10 for Adam's epsilon in the new record like we did before?

#176 [Bug] "Running the current record" doesn't work

#160 Hard time replicating on 8xH100

#145 GH200 support

#143 ModuleNotFoundError: No module named 'kernels'

#135 Can the evaluation system handle a large number of concurrent submissions?