Nov 15 Yesterday 22:32 Publish inference files at repository root Copy checkpoint-15258 configuration and matching model weights together with the repository tokenizer assets to the root. Training-only optimizer and trainer state files are omitted. 22:29 Organize root configuration files under config/ Move configuration JSON files into config/; keep README.md and .gitattributes at repository root. 05:48 Upload folder using huggingface_hub 02:40 Upload notebooks/rat77m_train_ready.ipynb with huggingface_hub 02:08 Upload notebooks/cells/cell13.py with huggingface_hub 02:08 Upload notebooks/cells/cell12_12__完了__final_20_を_Bucket_へ.py with huggingface_hub 02:07 Upload notebooks/cells/cell11_11__Trainer___Discord___Bucket___連続_grad.py with huggingface_hub 02:07 Upload notebooks/cells/cell10_10__学習前診断_1_step_.py with huggingface_hub 02:07 Upload notebooks/cells/cell09_9__Dataset_再マップ適用___Collator___train_eva.py with huggingface_hub 02:07 Upload notebooks/cells/cell08_8__データダウンロード____旧ID確認___再マップ前提チェック.py with huggingface_hub 02:07 Upload notebooks/cells/cell07_7__init_fixed_保存.py with huggingface_hub 02:07 Upload notebooks/cells/cell06_6__モデル初期化___健全性.py with huggingface_hub 02:07 Upload notebooks/cells/cell05_5__mT5_emb_移植_行_L2_0_5___v6実績_.py with huggingface_hub 02:07 Upload notebooks/cells/cell04_4__頻度リスト_語彙マップ構築_旧mT5_ID___新64k_ID_.py with huggingface_hub 02:07 Upload notebooks/cells/cell03_3__Discord___Bucket_ユーティリティ.py with huggingface_hub 02:07 Upload notebooks/cells/cell02_2__設定.py with huggingface_hub 02:07 Upload notebooks/cells/cell01_1__環境セットアップ.py with huggingface_hub 02:07 Upload notebooks/cells/cell00.py with huggingface_hub 02:07 Upload notebooks/cell11_trainer.py with huggingface_hub 02:07 Upload notebooks/rat77m_train_ready.ipynb with huggingface_hub 02:00 Upload notebooks/rat77m_train_ready.ipynb with huggingface_hub 01:51 init-fixed elementRMS=0.5 + id_remap 01:47 init-fixed elementRMS=0.5 + id_remap 01:45 Emb init: row L2=0.5 (v6). Element RMS=0.5 caused loss~89 01:40 init-fixed elementRMS=0.5 + id_remap 01:33 Add cell11 full source as .py for easy copy 01:32 TrainingArguments: version-safe kwargs (save_safetensors optional) 01:30 init-fixed elementRMS=0.5 + id_remap 01:25 Robust mT5 sentinel ID resolution (fix if convert_tokens_to_ids returns unk) 01:22 discord_notify reads os.environ each call; match working test headers 01:18 discord_notify: User-Agent + clearer 403 diagnostics 00:59 Fix SyntaxError in Training start discord_notify f-string 00:57 EvalSwapTrainer, disable_tqdm, Discord dumps full settings on start 00:53 Fix: Discord throttle, PrinterCallback off, eval split/mask, resume max step, unk handling 00:40 Send loss>35 and grad_norm spike warnings to Discord 00:31 Align abort streak=2, logging_steps=10, warmup-zone Discord alerts 00:25 Fix: mT5→64k ID remap, element RMS=0.5, LR=1e-4, pre-diag, streak abort 00:22 Critical fixes 1-6 + Discord@500 + Bucket replace@5000 + final-20 Nov 14 2 days ago 23:53 Use Trainer + BucketSyncCallback (hf buckets sync on_save) for G4 23:48 G4 Blackwell optimized + failure-mode guards (force) 23:47 tiny test file 23:47 Add G4-optimized notebook as new file; remove stale 23:47 G4 Blackwell optimized notebook (force replace) 23:47 Optimize for RTX PRO 6000 Blackwell (G4) + anti-regression guards 23:41 Add special_tokens_map.json 23:41 Add generation_config.json 23:41 Add tokenizer_config.json 23:41 Add config.json 23:41 Add training-ready Colab notebook (seq_len=1024, emb transplant, Bucket support) 23:38 Add training-ready Colab notebook + model/tokenizer configs (seq_len=1024)