coptic-htr-calamari

Coptic HTR (calamari)

Handwritten text recognition checkpoint for Coptic manuscript line crops. This is the stable checkpoint published by nomikos-project.

The Hub file is best.pt (calamari-pytorch-v1): CNN (40/60 filters) + 2 bidirectional LSTMs, line height 48, CTC charset size 39. Trained from scratch on the Coptic finetuning split (no separate Coptic pretraining corpus). A wider CNN (60/80) was worse on the same test set.

Intended use

• Transcribe already-segmented Coptic line images.
• Expert-in-the-loop HTR.

Out of scope: page-level layout, other scripts, or unsegmented folios.

How to use

weights_source: hf://nomikos-project/coptic-htr-calamari@stable

PYTHONPATH=. python scripts/hf/fetch_model.py coptic-calamari-v1 --registry-tag stable

from huggingface_hub import hf_hub_download
from src.inference.calamari import CalamariPredictor

path = hf_hub_download("nomikos-project/coptic-htr-calamari", "best.pt")
CalamariPredictor([path]).predict_files(["line.png"], output_dir="preds")

Training

Single-stage CTC training on Coptic GT (seed 1111, batch 32, 5 augmented copies + 1 original, LR 0.001, 150-epoch schedule):

• Data: data/processed/coptic/finetuning — 2258 / 282 / 283 train/val/test lines (line-level 80/10/10).
• Architecture: lstm_layers=2, conv0_filters=40, conv1_filters=60.
• W&B: coptic_training_lstm2 (run name coptic_training_lstm2).

best.pt is the lowest-val-CER checkpoint from that run.

Evaluation

Official test split (evaluation.split=test, 283 lines):

Compared on the same test set, the 60/80-filter variant scored CER 8.76% / WER 63.4% / exact match 38.9%, so this smaller CNN is the published stable model.

Files