AMIDDA 1.0 Line 4B
A foundation Vision Language Model for line-level Arabic handwritten text recognition (HTR), pre-trained on AMIDDA and developed at Calfa. Not designed for production.
Unlike the per-corpus specialised models that preceded it, this is a single generic model covering Maghrebi and Oriental Arabic manuscripts, Urdu lithographs, Persian, and modern Arabic handwriting.
| Base model | Qwen3.5-VL-4B-Instruct |
| Training data | AMIDDA, 53,420 training lines from 8 corpora |
| Method | LoRA fine-tuning, r = 64, adapter merged into the base weights |
| Task | Line-level HTR on pre-segmented images |
Presented at HIP 2026 @ ICDAR 2026. See the system report for data, training procedure, and full results.
Results
CER (%) on the AMIDDA test split (6,684 lines), per source corpus. Lower is better. The CTC row gives the per-corpus specialised RNN baselines for perspective; they are not a single generic model.
| Method | baybars | iskandar | khatt | muharaf | rasam1 | rasam2 | rasm | tarima | Overall |
|---|---|---|---|---|---|---|---|---|---|
| AMIDDA 1.0 Line 4B | 13.4 | 15.9 | 9.8 | 30.1 | 25.0 | 27.6 | 37.4 | 24.2 | 23.1 |
| QARI-OCR v0.3, zero-shot | 140.1 | 113.6 | 50.9 | 105.9 | 114.6 | 100.6 | 71.2 | 85.1 | 104.3 |
| Gemini 3 Flash, 25-shot ICL | 11.2 | 11.8 | 7.5 | 22.5 | 16.5 | 29.0 | 21.1 | 25.6 | 18.3 |
| CTC, one RNN per corpus | 7.8 | — | 6.1 | 9.5 | 3.6 | 4.9 | 8.9 | 3.9 | 6.4 |
Intended use
The model expects a cropped image of a single text line and returns its transcription. It is not a page-level OCR system: segmentation must be performed beforehand.
Transcriptions preserve the text as written in the original document, without normalization, modernization, or
editorial correction. Diacritics (tashkil) are transcribed only where the training guidelines record them, which
varies across source corpora; see the transcription_guidelines column of AMIDDA.
Dependencies
pip install "transformers>=5.9" torch accelerate pillow
transformers>=5.9 is required: earlier versions do not implement the qwen3_5 architecture.
The Qwen3.5 backbone alternates linear-attention and full-attention layers. Without the optimised kernels,
transformers silently falls back to reference PyTorch implementations that are correct but much slower. On CUDA,
install them:
pip install causal-conv1d flash-linear-attention
Usage
import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "calfa-ai/amidda1.0Line4B"
PROMPT = "Transcribe the handwritten text in this line image. Output ONLY the transcription."
processor = AutoProcessor.from_pretrained(model_id, max_pixels=401408)
model = AutoModelForImageTextToText.from_pretrained(
model_id, dtype=torch.bfloat16, device_map="auto"
)
model.eval()
image = Image.open("path/to/line_image.jpg").convert("RGB")
messages = [{
"role": "user",
"content": [
{"type": "image", "image": image},
{"type": "text", "text": PROMPT},
],
}]
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt",
enable_thinking=False,
).to(model.device)
with torch.no_grad():
generated = model.generate(**inputs, max_new_tokens=128, do_sample=False)
trimmed = generated[0][inputs["input_ids"].shape[1]:]
print(processor.decode(trimmed, skip_special_tokens=True).strip())
enable_thinking=False is required.
Usage with DocWorkflow
Training and evaluation were carried out with DocWorkflow, adapted to Arabic script. DocWorkflow handles ALTO XML input/output, line extraction, batching, and scoring in one pipeline.
git clone https://github.com/TheoMoins/DocWorkflow
cd DocWorkflow
pip install -e .
run_name: "amidda1.0Line4B"
output_dir: "results"
device: "cuda"
data:
test: "path/to/your/alto/data"
tasks:
htr:
type: VLMLineHTR
config:
model_name: "calfa-ai/amidda1.0Line4B"
device_map: "auto"
max_new_tokens: 128
line_batch_size: 8
max_pixels: 401408
prompt: >
Transcribe the handwritten text in this line image.
Output ONLY the transcription.
docworkflow -c amidda1.0Line4B.yml predict -t htr -d test
Related resources
All Arabic HTR resources are gathered in the Arabic HTR collection.
Training corpora: AMIDDA aggregates RASAM-1, RASAM-2, TARIMA, BAYBARS, ISKANDAR, KHATT, MUHARAF, and RASM (British Library, not redistributed).
Citation
@unpublished{vidalgorene:hal-05693582,
TITLE = {{Is a Generic Dataset and Foundation VLM for Arabic HTR Worth It? Lessons from AMIDDA}},
AUTHOR = {Vidal-Gor{\`e}ne, Chahan and Lucas, No{\"e}mie and Salah, Cl{\'e}ment and Decours-Perez, Ali{\'e}nor},
URL = {https://enc.hal.science/hal-05693582},
NOTE = {working paper or preprint},
YEAR = {2026},
MONTH = Jul,
KEYWORDS = {Arabic HTR ; Vision-Language Models ; In-Context Learning ; Data Bootstrapping ; DISTAM},
PDF = {https://enc.hal.science/hal-05693582v1/file/main.pdf},
HAL_ID = {hal-05693582},
HAL_VERSION = {v1},
}
License
The weights are released under Apache 2.0, following the Qwen3.5-VL-4B-Instruct base model. This covers the model only: the training data is mixed-license, and each AMIDDA sub-dataset carries its own terms. RASM is not redistributed.