Upload checkpoints/pretrain_step05252435.pt with huggingface_hub

Upload pretrain_step03817911.pt with huggingface_hub

Upload checkpoints/pretrain_step02804149.pt with huggingface_hub

Upload pretrain_step02550400.pt with huggingface_hub

Upload n.py with huggingface_hub

Upload checkpoints/pretrain_step02022699.pt with huggingface_hub

Upload checkpoints/pretrain_step01821614.pt with huggingface_hub

Upload checkpoints/pretrain_step01720930.pt with huggingface_hub

Upload pretrain_step01922101.pt with huggingface_hub

Upload pretrain_step01821614.pt with huggingface_hub

Upload checkpoints/pretrain_step01922101.pt with huggingface_hub

Upload pretrain_step01720930.pt with huggingface_hub

Add rank optimization test - validates 2x ratio is optimal

Add GQA attention module with checkpoint compatibility

Add experiments/README.md

Add experiments/infer_bench.py

Add experiments/final_showdown.py

Add experiments/joint_test.py

Add experiments/n_flex.py

Add experiments/n_ultra.py

Add experiments/n_heavy2.py

Add experiments/n_heavy.py

Upload checkpoints/pretrain_step01620126.pt with huggingface_hub

Add CHANGELOG documenting GradScaler fix

Fix GradScaler resume bug - wrapped scaler.load_state_dict() in try/except at line 512. Allows resuming from checkpoints saved without AMP.

Backup status.json

Backup latest.json

Backup script hf_upload.py

Backup script inference_api.py

Backup script rotating_log.py

Backup script patcher.py

Backup script stream_loader.py

Backup script n.py

Backup checkpoint pretrain_step01319973.pt

Backup checkpoint pretrain_step01034332.pt

Backup checkpoint pretrain_step00975327.pt

Checkpoint step 01178328 - 2026-01-13 13:39

Step 686090 checkpoint - 49hrs training

Add README with model details

Add tokenizer: tokenizer.json

Add tokenizer: special_tokens_map.json

Add tokenizer: tokenizer_config.json

Add hot-reload config template

Add checkpoint uploader script

Add dual rotating log (5000+2500 lines)

Add trainer: AR+SAT joint training, hot-config, HF auto-upload

Step 176907 - double chinchilla training (35.76B target)

Upload README.md with huggingface_hub

initial commit