All pull requests

Pull requests

• Reject data-loader resumes that would replay consumed tokens

  #931·salma2vec opened Oct 20, 2026·

  ·
  1

  1

• Preserve sampled token order during sequence-length warmup

  #930·salma2vec opened Oct 20, 2026·

  ·
  1

  1

• Fix skipped AdamW corruption, z-loss masking, and Ulysses masks

  #927·alimoradmohameddoweb opened Oct 19, 2026·

  ·
  6

• Mixed midtraining: build 2 rank batches ahead again

  #926·RustinS opened Oct 18, 2026·

  ·
  1

• Complete batch-rounded source mixtures while respecting source ratios and caps

  #922·AaravMehta-07 opened Oct 12, 2026·

  ·
  34

• Align melted list fields with their token sequences

  #921·AaravMehta-07 opened Oct 11, 2026·

  ·
  2

• Add OLMo Core <-> MaxText checkpoint conversion

  #920·CalebOuellette opened Oct 8, 2026Contributor·

  ·

• Add Dion3/NorDion2 optimizer support with LR and checkpoint LR-recovery fix

  #919·JenWei0312 opened Oct 7, 2026·

  ·

• [Do not merge yet] Mixed midtraining on OLMo 3.5 + #885: the OLMo-core side of scaling-ladders#83

  #916·jason718 opened Oct 6, 2026Contributor·

  ·
  

• Refresh README with official Olmo logo and capitalization

  #915·SchmittyThePirate opened Oct 6, 2026·

  ·

• [Multimodal 4/4] Stage-1 recipe: datasets, mixtures, and Molmo2-Stage1.py

  #913·jason718 opened Oct 5, 2026Contributor·

  ·
  

• [Multimodal 3/4] Multimodal data core: sequence builder, packing, collator, mixture loader

  #912·jason718 opened Oct 5, 2026Contributor·

  ·
  

• [Multimodal 2/4] Multimodal train module, PerGroupScheduler, weighted CE

  #911·jason718 opened Oct 5, 2026Contributor·

  ·
  17

  

• Fix Molmo2 stage-2 smoke recipe GPU count

  #907·wisoums opened Oct 4, 2026·

  ·

• Replay dropout masks under activation checkpointing (attention dropout included), and turn attention dropout off in eval

  #906·Mazukiri opened Oct 2, 2026·

  ·
  1

• docs: fix links to OLMo-2 32B data mixes and hybrid SFT script

  #896·pratikgx opened Sep 30, 2026·

  ·

• Register cached-path clients before hashing metadata sidecars

  #890·abhishekraok opened Sep 28, 2026Contributor·

  ·
  1

• Add opt-in Core interfaces for RL training and weight publication

  #888·no0p opened Sep 27, 2026·

  ·
  9

• Fix OLMoDDP Torch 2.13 training and preserve HF inference numerics

  #885·jacob-morrison opened Sep 26, 2026Contributor·

  ·
  9

• Consolidate DDP tests and fix tensor-parallel loss reduction

  #883·AkshitaB opened Sep 23, 2026Contributor·5/5·

  ·
  5

• Migrate scheduler aliases in training recipes

  #882·AkshitaB opened Sep 23, 2026Contributor·4/5·

  ·
  3

• Tune the FlexAttention backward with kernel_options (+12.7% Stage-2 useful TPS)

  #877·donovanclay opened Sep 21, 2026·

  ·

• Add mixed midtraining after vision alignment

  deprecated
  #874·RustinS opened Sep 19, 2026·

  ·
  6

  

• Log useful (non-pad) TPS and pack occupancy alongside padded TPS

  #870·donovanclay opened Sep 17, 2026·

  ·
  3

• Integrate native three-stage vision alignment

  deprecated
  #869·RustinS opened Sep 16, 2026·3 of 5 tasks·

  ·
  24