ept.git/ep_run/casc_bp_train.py, branch xiang Unnamed repository; edit this file 'description' to name the repository. Add portable 135M BP sweep handoff 2026-07-21T22:28:06+00:00 Oscar Wan oscarwan@oriong12.stanford.edu 2026-07-21T22:28:06+00:00 ad4622e42abf2a55a37c5e522a6f8302e2e95cfd Match the active EP configuration and support four-GPU BP candidates from collaborators' data locations. Co-authored-by: Cursor <cursoragent@cursor.com> 
Match the active EP configuration and support four-GPU BP candidates from collaborators' data locations.

Co-authored-by: Cursor <cursoragent@cursor.com>
save final ckpt at last step (fixed-eval protocol); fw72m_ride crown-3 launched (plain+ride governor); ride/BP seed campaign on GPU0 2026-07-17T05:44:11+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-17T05:44:11+00:00 835d327caa0a86d3b9fdc7b1008d717c5d4c5bed Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Stage-0 BP mirror arms: qup/qcomp flags in casc_bp_train; bp_qctl + 6 quant arms chained after qcomp (decision metric = Delta_EP vs Delta_BP) 2026-07-16T09:00:47+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-16T09:00:47+00:00 4b58cf8ba2e8e9ba167dc61dc4ed2e58dfdab4d8 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
wandb reorganized: per-regime team projects (ept-tinystories-42m / ept-fineweb-72m), auto project routing in trainers, split replay + reports 2026-07-14T18:04:38+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-14T18:04:38+00:00 836e72a707de4f8bd7679400cb40fe59464daca9 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
wandb -> team entity eqprop-llm-training (trainers + full-history replay script) 2026-07-14T17:53:22+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-14T17:53:22+00:00 8a9db1232a1bca01900f4113bb68666229384512 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Fix swallowed --wandb_run arg (replace-comment bug); wandb default-on intact 2026-07-13T18:56:44+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-13T18:56:44+00:00 e20ee6df345c5ba620b6af65391627775f44fe50 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
wandb ON by default in both trainers (project ept-cascade); BP twin gains --amp; fw72m_bp launched GPU1 2026-07-13T18:56:16+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-13T18:56:16+00:00 d15ae909dfb9c8870a79e6aea87e8085fda53460 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
Stage-2 data pipeline: FineWeb-Edu sample-10BT -> 32k BPE -> bins; --data flag on both trainers 2026-07-13T10:58:48+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-13T10:58:48+00:00 1d638edd83dc888cb392e902eb86ad57b22949a1 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
RESULT 10: epoch endpoints (BP 1.2509/1.2750 vs EP 1.4802, +0.22 erosion cost); GENERATION GATE PASSED (casc_gen.py, coherent stories, no-backprop 42.75M); stage1b improved-recipe pair launched (Muon cosine via build_hybrid) 2026-07-11T01:56:01+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-11T01:56:01+00:00 6d9fcd748df0836528b31f29717ac3bc8715d1dd Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn
OLMo2-standard architecture (user directive): norm-after-sublayer RMSNorm blocks, full-width QK-norm, RoPE(500k), SwiGLU, no-bias, untied head, final RMSNorm, 0.02 init, grouped wd, optional z-loss (plumbed through nudge+readout+gate for EP exactness); EP smoke 42.75M cos=1.0000 2026-07-10T13:24:03+00:00 Yuren Hao yurenh2@illinois.edu 2026-07-10T13:24:03+00:00 4e0d730765a8287cdd574dc39b8dafc16006b266 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn