frisket-ocr-student (Batch 01 @ 1k steps)
Early checkpoint of the Frisket OCR student fine-tuned from SmolVLM-256M on nvidia/OCR-Synthetic-Multilingual-v1 (Batch 01).
| Steps | 1000 |
| Train loss | 18.21 → 0.10 |
| Eval gate | not run yet |
| Hardware | RTX 3090 |
Not production-ready — mid-run snapshot for sharing / experiments.
from transformers import AutoProcessor, AutoModelForVision2Seq
repo = "mbx2/frisket-ocr-student-batch01-1k"
processor = AutoProcessor.from_pretrained(repo)
model = AutoModelForVision2Seq.from_pretrained(repo, torch_dtype="auto", device_map="auto")