Update train_sft.py with Pure English v2 anchors and domain mapping

Update train_dpo.py for calibrated DPO v2 training

Security: remove hardcoded tokens from code/build_dpo_dataset.py, use env/arg only

Security: remove hardcoded tokens from code/train_dpo.py, use env/arg only

Add SFT auxiliary loss (0.5x) and boost DPO learning rate to 1.5e-5 for 500M model

Upload latest DPO v1 alignment script: code/build_dpo_dataset.py

Upload latest DPO v1 alignment script: code/train_dpo.py

Upload code/train_sft.py with huggingface_hub

Delete code/build_sft_v2_chat_dataset.py with huggingface_hub

Delete code/build_sarus_sft_dataset_kaggle.py with huggingface_hub

Delete code/build_dpo_dataset.py with huggingface_hub

Upload code/config.py with huggingface_hub

Upload code/train_sft.py with huggingface_hub

Update code/tokenize_sft_v1_kaggle.py for SFT v1 21-domain curriculum

Update code/train_sft.py for SFT v1 21-domain curriculum

Only push final checkpoint to HF to save 10 min upload time

Upload both final and best checkpoints to Hugging Face

Update code/build_sarus_sft_dataset_kaggle.py with user enhancements

Update code/build_dpo_dataset.py with user enhancements

Update code/train_dpo.py with user enhancements

Update code/model.py with user enhancements

Update code/build_sft_v2_chat_dataset.py with user enhancements

Update code/train_sft.py with user enhancements

Upload code/build_sft_v2_chat_dataset.py with huggingface_hub

Upload code/build_sft_v2_chat_dataset.py with huggingface_hub

Upload code/train_sft.py with huggingface_hub

Update code/train_sft.py with chunked cross-entropy and H200 turbo speed optimizations

Update code/model.py with chunked cross-entropy and H200 turbo speed optimizations

Update code/model.py with chunked cross-entropy and H200 turbo speed optimizations

Update code/train_sft.py with chunked cross-entropy and H200 turbo speed optimizations

Update code/model.py with chunked cross-entropy and H200 turbo speed optimizations

Enable pure FlashAttention-2 & 8 async dataloader workers

Update H200 ultra-tier settings

Tune VRAM memory: delete logits before backward, safe micro_batch 4 for 32GB GPUs

Fix argparse formatting, add VRAM safety guard & local shard reuse in train_sft.py

Add modular phases (raw, check, tokenize) and automated forensic audit

Update train_sft.py: DOMAIN_NAMES_V1 9 domains without safety

Update builder script: 9 domains, integrate user identity and greetings, remove safety

Fix D1 UltraChat greetings & D2 NCERT None handling in build_sarus_sft_dataset_kaggle.py

Add Sarus SFT v1 code/train_sft.py

Add Sarus SFT v1 code/build_sarus_sft_dataset_kaggle.py

Upload code/train_sft.py with huggingface_hub

Upload code/train_sft.py with huggingface_hub

Fix audit bugs in code/model.py (tail drop, loss path, dynamic domains, repetition penalty)

Fix audit bugs in code/train_sft.py (tail drop, loss path, dynamic domains, repetition penalty)

Update code/train_sft.py with audited production fixes

Update code/config.py with audited production fixes

Update code/model.py with audited production fixes

Clean obsolete model files (batch 1)

Update train_sft.py with SFT v26 support