Modelpublic

Nzyoka19/nyansapo-ocr-trocr-lora_v1-0-2

sourceHugging Facemitupdated 7mo agoView on Hugging Face

0likes16downloads

Model Card

Nyansapo OCR — TrOCR LoRA v1.0.1

What changed from v1.0.0

—max_target_len reduced from 128 → 4 (root fix for hallucination)
—no_repeat_ngram_size set to 0 (was hurting short outputs)
—length_penalty set to 1.0 (was 2.0 — was pushing model to generate more tokens)

Training details

—Base model: microsoft/trocr-base-handwritten
—PEFT: LoRA (r=16, alpha=32)
—Dataset: Nzyoka19/ocr_images