Issues

• H156 run log: capped science traces swap on Olmo 3 7B (TPU)

  agent-update-log
  allenai/open-instruct#2142 · Filter by author abhishekraok opened on Nov 2, 2026

• H154 run log: Kimi K3 IF/chat rewrite on Olmo 3 7B (TPU)

  agent-update-log
  allenai/open-instruct#2139 · Filter by author abhishekraok opened on Nov 1, 2026
  2

• H161 run log: hero C12 SFT, two epochs on base B1 (TPU)

  agent-update-log
  allenai/open-instruct#2130 · Filter by author abhishekraok opened on Oct 31, 2026
  3

• Olmo 3.5 TPU converters and MILES presets hard-code the current layout (CYCLE = 8)

  allenai/open-instruct#2123 · Filter by author abhishekraok opened on Oct 30, 2026

• MILES code reward: HTTP 500 from the code service scores 0 with no retry

  allenai/open-instruct#2122 · Filter by author abhishekraok opened on Oct 30, 2026

• H158 run log: C12 SFT on midtraining checkpoint 3 (alt vision alignment) vs B2 (TPU)

  agent-update-log
  allenai/open-instruct#2119 · Filter by author abhishekraok opened on Oct 30, 2026
  1

• Dataset caches for other local sources (directories, .json, .csv, .jsonl.gz) still key on path, not content

  allenai/open-instruct#2104 · Filter by author SUPR1M3 opened on Oct 29, 2026

• H152 tmax-big-b2b1: tmax-sft-big continuation on H134's b2-c11 / b1-c11 on TPU (run log)

  agent-update-log
  allenai/open-instruct#2100 · Filter by author pdasigi opened on Oct 29, 2026
  14

• H147 rl-b2-c11: RL (H066 recipe) on C11 SFT from the B2 midtraining checkpoint — update log

  agent-update-log
  allenai/open-instruct#2099 · Filter by author abhishekraok opened on Oct 28, 2026
  3

• DeepSpeed bf16 casts the rotary inv_freq buffers to bf16 during accelerator.prepare

  allenai/open-instruct#2098 · Filter by author armancohan opened on Oct 28, 2026

• OLMo 3 SFT under transformers 5.4.0 applies YaRN to sliding-window layers, unlike vLLM and OLMo-core

  allenai/open-instruct#2097 · Filter by author armancohan opened on Oct 28, 2026

• H148: The olmo-eval sglang provider reproduces vLLM eval scores on an OLMo 3 7B Think SFT checkpoint

  agent-update-log
  allenai/open-instruct#2094 · Filter by author armancohan opened on Oct 28, 2026
  7

• H120 run log: C11 vs C10 through the Olmo 3 Think RL recipe (4T EMO)

  agent-update-log
  allenai/open-instruct#2092 · Filter by author abhishekraok opened on Oct 28, 2026
  5

• H119 c11-decon-14t: does dropping the flagged-pair Dolci rows cost C11 anything? (run log)

  agent-update-log
  allenai/open-instruct#2085 · Filter by author abhishekraok opened on Oct 27, 2026

• H135: Adding 100K DeepSeek-V4.1-Flash traces on new TextbookReasoning questions to H055's 200K Dolci-Think control improves science

  agent-update-log
  allenai/open-instruct#2084 · Filter by author armancohan opened on Oct 27, 2026

• H138: Generated science-document CPT before matched SFT improves science reasoning

  agent-update-log
  allenai/open-instruct#2083 · Filter by author armancohan opened on Oct 27, 2026

• H129: Spark-234K science problems re-traced with DeepSeek-V4.1-Flash, added to H055's Dolci-Think control, improve science reasoning

  agent-update-log
  allenai/open-instruct#2082 · Filter by author armancohan opened on Oct 27, 2026

• H134 mm-ckpt-sft: C11 SFT on TPU from three midtraining checkpoints (run log)

  agent-update-log
  allenai/open-instruct#2081 · Filter by author abhishekraok opened on Oct 27, 2026
  5

• Olmo 3.5 SFT H045: thinking runs to max_tokens without closing </think> (no final answer on 18% of turns)

  olmo voice issues
  allenai/open-instruct#2080 · Filter by author ginaw-ai2 opened on Oct 27, 2026
  2

  

• H133: Adding 100K Nemotron Science-v2 synthetic RQA traces to H055's 200K Dolci-Think control improves science

  agent-update-log
  allenai/open-instruct#2078 · Filter by author armancohan opened on Oct 26, 2026

• TPU SFT resume: the R2-emulation data shim re-reads every trained row (75–120 min idle on a big slice)

  allenai/open-instruct#2074 · Filter by author abhishekraok opened on Oct 25, 2026

• H128: Adding 50K Nemotron IF-Chat-v3 rows to a 100K IF-and-chat Dolci-Think control improves instruction following and chat

  agent-update-log
  allenai/open-instruct#2073 · Filter by author armancohan opened on Oct 25, 2026

• H126: Adding about 100K DeepMath-103K R1 traces to H055's 200K Dolci-Think control improves math

  agent-update-log
  allenai/open-instruct#2070 · Filter by author armancohan opened on Oct 24, 2026

• H113: Olmo 3 7B from base on TPU, Nemotron-Science-v1 traces added to H055's Dolci-Think control vs the same control on TPU

  agent-update-log
  allenai/open-instruct#2065 · Filter by author armancohan opened on Oct 23, 2026

• H103 run log: Olmo 3.5 small, 4x batch at peak LR 1e-4 on TPU (seed 1 on one v4 8x8x16)

  agent-update-log
  allenai/open-instruct#2064 · Filter by author abhishekraok opened on Oct 23, 2026
  4