From 78c24e20145736fb48737e480bf446a600baf6a0 Mon Sep 17 00:00:00 2001 From: spokane-way Date: Thu, 19 Mar 2026 14:04:16 -0700 Subject: New SOTA attempt (#52) Co-authored-by: spokane-way --- .../2026-03-19_TrainingOptSeq4096/submission.json | 11 +++++++++++ 1 file changed, 11 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-19_TrainingOptSeq4096/submission.json (limited to 'records/track_10min_16mb/2026-03-19_TrainingOptSeq4096/submission.json') diff --git a/records/track_10min_16mb/2026-03-19_TrainingOptSeq4096/submission.json b/records/track_10min_16mb/2026-03-19_TrainingOptSeq4096/submission.json new file mode 100644 index 0000000..71d9851 --- /dev/null +++ b/records/track_10min_16mb/2026-03-19_TrainingOptSeq4096/submission.json @@ -0,0 +1,11 @@ +{ + "author": "Spokane Way", + "github_id": "spokane-way", + "name": "Training Opt Seq4096 v1", + "blurb": "SP-1024 9x512 KV4 run at TRAIN_SEQ_LEN=4096 with aggressively tuned Muon optimizer: momentum 0.99, lower LR (0.020/0.020/0.030), 3/4 batch (393K tokens), warmdown 3000 steps, and extended momentum warmup (1500 steps from 0.92). Combines long-context training with training optimization to beat the naive baseline by 0.023 BPB.", + "date": "2026-03-19T04:28:00Z", + "val_loss": 2.02857127, + "val_bpb": 1.20143417, + "bytes_total": 15868326, + "bytes_code": 47759 +} -- cgit v1.2.3