armenian-htr-calamari

nomikos-project/armenian-htr-calamari

armenian-htr-calamari is a image-to-text model from nomikos-project. Use it when you need a caption or text from an image. It is set up for calamari.

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

Downloads · 30 days

31% of all-time downloads

All-time downloads

Public

Repo size

30.1 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.onnx10.1 MB · 50%

At a glance

Task: Image-to-Text
Library: calamari
Access: Public
Created: Sep 4, 2026
Updated: Sep 7, 2026
SHA: ed5d7423

Model card

From the Hugging Face model README

Armenian HTR (calamari)

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

Field	Value
registry model id	`armenian-calamari-v1`
registry tag	`stable`
script	`armenian`
architecture	`calamari`
model version	`v1`
task	`transcribe`
weights source	`hf://nomikos-project/armenian-htr-calamari@stable`

The Hub file is best.pt (calamari-pytorch-v1): CNN + 2 bidirectional LSTMs, line height 48, CTC charset size 96. Some training YAML files originally said lstm_layers: 3; that setting was not applied, and the saved weights are 2 BiLSTMs.

Intended use

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

Out of scope: page-level OCR or other scripts.

How to use

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

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

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

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

Training

Cleaned Armenian GT (seed 1111, batch 32, 5 augmented copies + 1 original):

Pretrain 150 epochs, LR 0.001 on data/processed/armenian/pretraining (1,807 / 297 / 282). W&B: v21t1dvl.
Finetune 100 epochs, LR 0.001 from that pretrain on data/processed/armenian/finetuning (1,141 / 119 / 120). W&B: jz4kk1wx.

best.pt is selected on validation CER.

Evaluation

Scored from this Hub best.pt with python -m src.evaluate.calamari on data/processed/armenian/finetuning.

Metric	Val (119 lines)	Test (120 lines)
CER	0.092	0.072
WER	0.440	0.340
Exact match	0.319	0.458
SROIE F1	0.588	0.701

best.pt is selected on val CER (0.092). W&B eval/* during training is that same val split, not test.

Limitations

Pretraining set is much smaller than Greek/Syriac.
Short lines have higher CER than typical 13–20 character lines.
Human correction remains necessary.