Issues

• H217 clean-plus-ota-b2: cleaned tmax-sft-big + OpenThoughts-Agent-SFT-100K (no issue-tasks) vs H197 clean-b2 (update log)

  agent-update-log
  allenai/open-instruct#2249 · Filter by author pdasigi opened on Nov 12, 2026
  9

• H214 run log: Dolci-Think with judge-filtered Qwen3.8-27B model-spec completions (model B)

  agent-update-log
  allenai/open-instruct#2247 · Filter by author gregorybchris opened on Nov 12, 2026
  3

• numpy SFT conversion: resume crashes with IndexError when the collect loop had finished

  allenai/open-instruct#2245 · Filter by author abhishekraok opened on Nov 12, 2026

• H212 run log: Dolci-Think with Qwen3.8-27B model-spec completions swapped in

  agent-update-log
  allenai/open-instruct#2240 · Filter by author gregorybchris opened on Nov 12, 2026
  11

• Olmo 3 7B Instruct: exact SFT → DPO → RL checkpoint lineage

  allenai/open-instruct#2219 · Filter by author quick910516-byte opened on Nov 9, 2026

• ray_node_setup.sh: a worker that never reaches its head exits 0, so a misconfigured job reports success

  bug
  allenai/open-instruct#2205 · Filter by author abhishekraok opened on Nov 7, 2026
  1

• H199: H161 Dolci control versus expanded RL dataset pools

  agent-update-log
  allenai/open-instruct#2201 · Filter by author no0p opened on Nov 7, 2026
  4

• Audit Dolci stdio targets: non-text test values and function-call problems routed as stdio

  allenai/open-instruct#2200 · Filter by author no0p opened on Nov 7, 2026

• Regenerate WildGuardMix / WildJailbreak / CoCoNot without the hidden safety system prompt.

  SFT
  allenai/open-instruct#2196 · Filter by author abhishekraok opened on Nov 6, 2026
  

• H190 c12-txt (text-only twin of B2, C12 SFT on TPU): run log

  agent-update-log
  allenai/open-instruct#2190 · Filter by author abhishekraok opened on Nov 6, 2026
  3

• Olmo Hybrid Instruct SFT/DPO: are safety-suite numbers available?

  allenai/open-instruct#2180 · Filter by author gloriaowowo opened on Nov 4, 2026
  1

• H187 update log: RL loss averaging (token vs sequence vs constant) on C10

  agent-update-log
  allenai/open-instruct#2173 · Filter by author farhatkevin opened on Nov 4, 2026
  11

• Curate the changelog around user impact and remove path-based enforcement

  allenai/open-instruct#2169 · Filter by author abhishekraok opened on Nov 3, 2026

• 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
  4

• 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

• 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

• 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

• 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