AMIDDA — A Dataset for Generic Arabic ATR model

AMIDDA (Arabic Manuscript Images for Digital Deciphering and Analysis) is a combined dataset aggregating line-level ground-truth from eight HTR datasets covering Arabic (Maghrebi & Oriental scripts), Urdu, Persian, and modern Arabic handwriting.

Stats

Sub-datasets	8
Lines (total)	66,779
Train / Val / Test	53,420 / 6,675 / 6,684

Sub-datasets

Dataset	Script	Source	Lines
RASAM-1	Arabic (Maghrebi)	BULAC	7,331
RASAM-2	Arabic (Maghrebi)	BULAC	3,720
TARIMA	Arabic (Maghrebi)	BULAC / BnF / MMSH	2,285
BAYBARS	Arabic (Oriental)	—	15,567
ISKANDAR	Arabic (Oriental)	BnF / Gotha / Leeds	5,221
RASM	Arabic (Oriental)	British Library (not redistributed)	1,488
KHATT	Arabic (modern cursive)	KHATT v1.0	6,673
MUHARAF	Arabic (modern cursive)	muharaf-public	24,494

Each row carries two source-related columns:

• dataset — exact source: rasam1, rasam2, tarima, baybars, iskandar, rasm, khatt, muharaf
• transcription_guidelines — guideline followed during transcription:
  • rasam — RASAM-1, RASAM-2, TARIMA, BAYBARS, ISKANDAR
  • rasm — RASM
  • khatt — KHATT
  • muharaf — MUHARAF

Only the rasam guidelines have been formally specified and verified. The other datasets are nonetheless highly compatible with them: all sources preserve the text as written in the original document, without normalization, modernization, or editorial correction.

Usage

from datasets import load_dataset

ds = load_dataset("calfa-ai/AMIDDA")

# Filter by source dataset
baybars_only = ds["train"].filter(lambda x: x["dataset"] == "baybars")

# Filter by transcription guideline
rasam_only   = ds["train"].filter(lambda x: x["transcription_guidelines"] == "rasam")

Results on this dataset

CER (%) per source dataset across the six strategies, one row per model variant, with third-party scores for perspective. "—" means not evaluated (Gemini 3.5 Flash: partial run on three corpora).

Method	baybars	iskandar	khatt	muharaf	rasam1	rasam2	rasm	tarima	Overall
(i) VLM API + line-level ICL — 280-line subset (35 per corpus), best shot count per model
Gemini 3 Flash, 25 shots	11.2	11.8	7.5	22.5	16.5	29.0	21.1	25.6	18.3
Gemini 3.5 Flash, 25 shots	9.6	11.0	4.8	—	—	—	—	—	9.9
Gemma 4 26B IT, 5 shots	22.9	20.7	18.8	42.5	33.1	40.8	32.2	39.3	31.6
Mistral Large 2512, 5 shots	56.3	62.9	65.0	70.3	70.4	67.8	73.7	63.3	66.7
Mistral Small 2603, 5 shots	68.7	66.5	62.6	75.3	72.3	75.2	77.6	70.3	71.1
Qwen 2.5 VL 72B, 5 shots	22.9	28.1	21.1	49.0	43.4	43.7	42.0	53.9	38.7
Qwen 3 VL 8B, zero-shot	53.9	47.8	104.7	89.7	96.4	108.9	57.5	82.7	83.4
Qwen 3.5 Flash, zero-shot	32.3	26.9	41.0	55.2	55.9	56.4	40.2	55.6	46.8
Qwen 3.6 Flash, zero-shot	31.9	38.9	36.6	55.2	60.6	60.0	41.8	59.4	49.4
(ii) VLM API + page-level context — 60-line subset (10 per corpus), best mode per model (KHATT and Muharaf excluded)
Gemini 3 Flash, ctx = text	23.2	11.8	—	—	26.4	32.8	10.6	26.0	21.8
Gemma 4 26B IT, ctx = text	37.7	19.8	—	—	45.4	45.2	21.3	46.6	36.0
Mistral Large 2512, ctx = img_text	72.8	68.4	—	—	75.7	70.6	61.4	69.5	69.7
Qwen 2.5 VL 72B, ctx = text	45.8	25.1	—	—	63.2	52.0	27.0	49.7	43.8
Qwen 3.6 Flash, ctx = text	52.4	36.3	—	—	67.1	57.0	63.8	76.6	58.8
(iii) Local zero-shot OCR-VLM — full test split
GLM-OCR, zero-shot	199.4	189.1	247.6	295.5	383.9	329.4	290.6	359.7	286.1
Chandra-OCR-2, zero-shot	496.4	553.8	657.9	1898.4	1536.8	1632.2	1056.6	1877.0	1346.2
QARI-OCR v0.3, zero-shot	140.1	113.6	50.9	105.9	114.6	100.6	71.2	85.1	104.3
GLM-OCR, LoRA fine-tuned	120.3	137.8	60.2	107.2	175.3	178.3	230.0	159.2	125.9
(iv) LoRA fine-tuning of Qwen3.5-4B — full test split
Qwen3.5-4B, r = 64	13.4	15.9	9.8	30.1	25.0	27.6	37.4	24.2	23.1
Qwen3.5-4B, r = 128	13.5	15.9	8.1	31.1	22.3	29.2	34.3	31.2	23.2
(v) MoE post-correction — 80-line subset, judge output
baseline (Qwen3.5-4B FT)	18.9	14.2	11.6	18.3	21.0	23.3	19.9	25.6	19.5
Gemini / Gemini / GPT-5.1	14.0	10.9	12.9	15.1	14.3	21.4	13.8	20.7	15.8
Gemini / Qwen3-VL-32B / GPT-5.1	15.6	10.4	11.0	16.0	17.6	21.7	21.0	20.7	17.2
Gemini / Qwen3.5-35B / Mistral-L	29.7	17.2	14.3	17.7	15.5	29.8	29.0	26.6	22.3
(vi) RNN (CTC), specialised per corpus, already existing — per-corpus test splits inherited by AMIDDA
CTC, one RNN per corpus (CALFA CRNN model, see RASAM paper)	7.8	—	6.1	9.5	3.6	4.9	8.9	3.9	6.4
Third-party baselines (reported on their own test splits)
HATFormer	—	—	15.4	8.6	—	—	—	—	—
Qalam	—	—	10.4 †	—	—	—	—	—	—

Source: AMIDDA -- Is a Generic Dataset and Foundation VLM for Arabic HTR Worth It? Lessons from AMIDDA hal-05693582

Citation

Presented at HIP 2026 @ ICDAR 2026 (proceedings to come).

Preprint:

@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

Mixed — see each sub-dataset for its own license. Refer to the source repositories.