Repository Issue Activity (beta)

#1895 Dolci-Think SFT anchor on the multi-trillion-token KDA checkpoint: re-establish the #1853 anchor and measure the pretraining-token effect

#1903 Non-termination on the Dolci-Think SFT anchors: SFT-budget and proxy-vs-hero 2x2 on matched Math-500 (ledger H012/H013)

#1902 Olmo Identity Data

#1898 Dolci-Think SFT anchor on the hero small PT-EMO-only checkpoint: paired arm for the EMO-in-pretraining question

#1899 Mismatch between allenai/Olmo-3-7B-RL-Zero-Code chat_template.jinja and the training template displayed in commit 70b0472af73bd6428e14c5d73abebbabaaeb4997

#1880 Anneal-matched 0.5-epoch anchor for the Olmo 3.5 SFT mixture search (KDA 1.2B, Dolci-Think, seq 65536)

#1869 Add a think token

#1896 transformers 5 loads Dolma 2 tokenizer-only repos with the GPT-2 pre-tokenizer, not the cl100k regex in tokenizer.json: SFT segmentation differs from pretraining (+9.8% tokens, 72% of Dolci-Think messages)

#1888 Dense Olmo 3 7B Dolci-Think 1-epoch SFT as an infra check against the released Think-SFT

#1875 Replicate Olmo 3 Think SFT on the KDA proxy: 2 epochs of Dolci-Think + paper-comparable evals

#1894 Verify evals on Papers with Code

#1891 Fix terminal evaluation recovery and SFT data filtering for the KDA checkpoint

#1857 MoE think DPO on the LC KDA 1.2B: Dolci-Think-DPO-7B via the olmo-core DPO path

#1853 Auto-research: SFT data mixture search on a fast proxy for Olmo 3.5

#1854 LC MoE think SFT baseline: 0.5 epoch Dolci-Think on the long-context KDA 1.2B

#1859 1-epoch Dolci-Think SFT at seq 65536 on the LC KDA MoE (checkpoint for agentic SFT)

#1882 Hub think-dev tokenizer breaks SFT label derivation; #1879 does not reach it

#1858 RL on the KDA MoE proxy (moved to internal tracker)

#1804 Thinker chat templates cannot be used for SFT: every row's labels are undecidable

#1805 `--chat_template_name` silently accepts unrecognised names and trains with a different template

#1855 64K infra gate for agent-trajectory SFT on the LC KDA MoE 1.2B

#1849 MoE + Dolci-Think SFT baseline: seq 16384, reasoning-general math gains where instruct SFT gave 0.0000

#1850 Finalize Post training evaluation suite

#1852 Mix science SFT dataset.

#1851 Mix multimodal SFT dataset.