Update weights: 10B token continue-pretraining on finepdfs eng_Latn (val ppl 36.8 → 32.3)

Fix weights: strip _orig_mod. prefix (torch.compile artifact) so model loads without workarounds

Upload folder using huggingface_hub