All pull requests

Pull requests

• Fix Olmo 3 RoPE: require transformers 5.17, flatten rope_parameters for vLLM

  #2077·edwin-huang opened Oct 26, 2026·

  ·

• Fix FA4 AuxData JIT argument warnings in the MILES image

  #2075·no0p opened Oct 25, 2026Collaborator·

  ·
  3

• Add --extra_numpy_dirs to OLMo-core SFT

  #2035·abhishekraok opened Oct 21, 2026Collaborator·

  ·
  6

• olmo-core SFT: document boundaries from row metadata by default

  #2028·abhishekraok opened Oct 19, 2026Collaborator·

  ·
  5

• Bump OLMo-core to 6ac500d8 (v2.6.0) for metadata document boundaries

  #2027·abhishekraok opened Oct 19, 2026Collaborator·

  ·
  3

• Integrate MILES NGU rejected-group handling and retry metrics

  #2025·mnoukhov opened Oct 19, 2026Contributor·

  ·
  5

• Add a standalone Qwen3 Deepscaler benchmark for MILES

  #2024·mnoukhov opened Oct 19, 2026Contributor·

  ·
  2

• Add enable_thinking to the production olmo35 chat template

  #2022·farhatkevin opened Oct 19, 2026Collaborator·

  ·
  5

• [experiment, not for merge] H087 IF gate-yes reward on ea00aceb

  #2013·abhishekraok opened Oct 17, 2026Collaborator·

  ·

• Let Codex load the repo skills; fix stale agent instructions

  #2006·abhishekraok opened Oct 16, 2026Collaborator·

  ·
  6

• Reconcile MILES verifier-GRPO reward evidence on resume

  #2003·farhatkevin opened Oct 16, 2026Collaborator·

  ·
  10

• Add a pinned report-only decon receipt adapter

  #1989·abhishekraok opened Oct 16, 2026Collaborator·

  ·
  6

• Add --min_runtime and --auto_resume to mason

  #1985·farhatkevin opened Oct 15, 2026Collaborator·

  ·
  11

• Add experimental Megatron verifier GRPO with a DPPO trust-region loss

  #1975·farhatkevin opened Oct 14, 2026Collaborator·

  ·
  17

• Remove stale and redundant README content

  #1971·abhishekraok opened Oct 14, 2026Collaborator·

  ·
  16

• Add opt-in SFT filter for literal added-token strings in content

  #1970·abhishekraok opened Oct 14, 2026Collaborator·

  ·
  18

• docs: add Simplified Chinese translation (README.zh.md)

  #1960·JasonYeYuhe opened Oct 12, 2026·

  ·
  5

• TPU: DPO on MaxText, and three SFT data fixes that make the Olmo 3 7B repro faithful

  #1953·abhishekraok opened Oct 11, 2026Collaborator·

  ·

• Add an opt-in guard that drops SFT rows spelling out special tokens

  #1942·abhishekraok opened Oct 7, 2026Collaborator·

  ·
  12

• Make local DPO debug runs work outside Ai2 infra

  #1941·stanleyIl opened Oct 6, 2026·

  ·
  2

• Normalize legacy SFT loss across accumulation windows and ranks

  #1939·abhishekraok opened Oct 5, 2026Collaborator·

  ·
  11

• H031: seed </think>'s lm_head row from its first piece (experiment branch)

  #1931·abhishekraok opened Oct 3, 2026Collaborator·

  ·
  3

• Make the Beaker CLI helpers fail soft, and stop looking datasets up twice

  #1928·VishnuR23 opened Oct 1, 2026Contributor·

  ·

• Compute GRPO advantages before masking truncated completions

  #1920·farhatkevin opened Sep 26, 2026Collaborator·

  ·
  5

• Clamp CodeVerifier perf-penalty multiplier at 0 to avoid negative rewards

  #1918·linhongyu510 opened Sep 24, 2026·

  ·
  1