Teklia/Qwen2.5-VL-7B-DAI-CReTDHI-RecordGold-ATR

QWEN2.5-VL - French Parish Register Model

Introduction

This version of QWEN2.5-VL-7B is specialized for HTR on French parish records from the 16th-18th centuries. Ref: https://redmine.teklia.com/issues/11177

Training

The model is a QWEN2.5-VL-7B-Instruct fine-tuned on French parish records using LoRA.

It was fine-tuned in two steps:

1. First, it was fine-tuned on 30k crowdsourced records (annotations of variable quality) → DAI-CReTDHI-RecordGeneanet-ATR
2. It was then fine-tuned on 7k records (high quality annotation) → DAI-CReTDHI-RecordGold-ATR

Parameters:

• Image width: 1500 pixels (max)
• LoRa rank: 8
• LoRa alpha: 32
• Epochs: 10 (about 4k steps)

Wandb: https://wandb.ai/starride-teklia/uncategorized/runs/bjv98gfo

Evaluation

It achieves the following results:

Overall evaluation

Model	CER (%)	WER (%)	NLS (%)
Qwen-2.5 7B VL	44.10	73.77	56.67
Qwen-3 8B VL	62.06	85.42	39.20
Qwen-3.5 9B VL	44.56	67.58	56.70
Churro-3B	22.24	36.26	79.87
Gemma4	55.49	85.75	46.21
Claude Opus 4.5	20.15	39.58	80.78
Mistral OCR	41.35	76.48	61.02
Gemini 3.0 Pro	16.31	33.47	84.25
Qwen-2.5 7-B VL (fine-tuned on gold data)	14.89	29.27	90.91
Qwen-2.5 7-B VL (fine-tuned on crowdsourced data)	11.41	24.24	-
Qwen-2.5 7-B VL (fine-tuned on crowdsourced and gold data)	9.24	21.25

Per period evaluation

Model	Period	CER (%)	WER (%)	NLS (%)	N images
Qwen-2.5 7B VL	< 1600	76.74	99.78	24.38	13
	1600 - 1650	57.33	89.14	43.19	169
	1650 - 1700	37.98	69.69	62.76	162
	1700 - 1750	39.59	68.48	60.91	210
	1750 - 1800	47.02	70.64	53.39	204
Qwen-3 8B VL	< 1600	99.85	100.0	0.15	13
	1600 - 1650	81.82	98.88	18.9	169
	1650 - 1700	52.25	78.67	48.89	162
	1700 - 1750	59.5	82.17	42.06	210
	1750 - 1800	59.0	80.92	42.32	204
Qwen-3.5 9B VL	< 1600	89.96	98.05	13.03	13
	1600 - 1650	66.58	87.65	34.8	169
	1650 - 1700	39.4	63.79	61.84	162
	1700 - 1750	31.84	56.44	68.81	210
	1750 - 1800	39.7	59.89	61.13	204
Churro-3B	< 1600	36.12	55.1	66.38	13
	1600 - 1650	23.23	40.72	77.41	169
	1650 - 1700	20.04	35.47	80.9	162
	1700 - 1750	21.41	32.51	79.15	210
	1750 - 1800	31.75	41.16	69.57	204
Gemma4	< 1600	115.4	137.96	22.34	13
	1600 - 1650	124.85	137.53	22.7	169
	1650 - 1700	77.2	106.59	40.97	162
	1700 - 1750	63.26	101.87	44.84	210
	1750 - 1800	70.96	99.36	43.95	204
Claude Opus 4.5	< 1600	68.5	104.99	46.62	13
	1600 - 1650	34.33	60.6	68.1	169
	1650 - 1700	21.84	42.87	79.4	162
	1700 - 1750	19.95	38.13	80.96	210
	1750 - 1800	14.17	31.39	86.24	204
Mistral OCR	< 1600	95.9	145.55	23.19	13
	1600 - 1650	65.71	107.68	38.86	169
	1650 - 1700	44.73	83.86	59.43	162
	1700 - 1750	38.31	72.84	63.45	210
	1750 - 1800	33.53	65.03	67.9	204
Gemini 3.0	< 1600	53.91	84.82	50.87	13
	1600 - 1650	28.65	54.02	72.17	169
	1650 - 1700	19.27	37.74	81.41	162
	1700 - 1750	14.17	29.63	86.33	210
	1750 - 1800	12.07	27.13	88.33	204
Qwen-2.5 7-B VL (fine-tuned on crowdsourced and gold data)	< 1600	31.2	50.11	71.95	13
	1600 - 1650	15.77	30.5	84.79	169
	1650 - 1700	8.91	20.45	91.26	162
	1700 - 1750	8.82	20.18	91.32	210
	1750 - 1800	7.19	18.88	92.87	204

Usage

Here we show a code snippet to show you how to use the model with transformers and qwen_vl_utils:

• Prediction script

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

# Load QWEN
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Teklia/Qwen2.5-VL-7B-DAI-CReTDHI-RecordGold-ATR",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("Teklia/Qwen2.5-VL-7B-DAI-CReTDHI-RecordGold-ATR", use_fast=True)

# Prompt
messages = [
    {
        "role": "system",
        "content": [
           {
               "type": "text",
               "text": "Tu es un assistant archiviste. Tu dois lire des actes issus de registres paroissiaux français, du 16è au 18è siècle. Extrais le texte de la marge, du corps de l'acte, et éventuellement les signatures."
           }
        ]
    },
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://europe.iiif.teklia.com/iiif/2/geneanet%2FArdennes_BMS%2F382706%2F00056.jpg/1252,102,1133,692/full/0/default.jpg"
            },
            {
                "type": "text",
                "text": "Extrais le texte de ce document."
            },
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=1024)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)[0]
print(output_text)

• Output

L'an mil sept cent quatre vingt six le dix septième jour du mois de mars est décédé à Frainvrit de cette paroisse Lambert Joseph Henrot âgé de trente huit ans, natif de Gumiay fils de Lambert Henrot et de Marie Thérèse Gervant lequel a été inhumé le lendemain au cimetière de cette ditte paroisse avec les ceremonies ordinaires par nous prêtre vicaire de cette ville, en présence de Vincent Parrot fleure de cette église, et de Piacre Blondeau, habitant de cette ville, lesquels ont signé avec nous.

Blondeau Parrot Berin vicaire

Citation

To cite the original QWEN2.5-VL model:

@misc{qwen2.5-VL,
    title = {Qwen2.5-VL},
    url = {https://qwenlm.github.io/blog/qwen2.5-vl/},
    author = {Qwen Team},
    month = {January},
    year = {2025}
}