Teklia/Qwen2.5-VL-7B-DAI-CReTDHI-IndexCardsNested-KIE

Qwen2.5-VL-7B-Instruct Index Cards Nested

Introduction

This version of QWEN2.5-VL-7B is specialized for document parsing on French index cards. It was fine-tuned as part of the DAI-CReTDHI project.

Training

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

Parameters:

• Image width: 800 pixels
• LoRa rank: 8
• LoRa alpha: 32
• Epochs: 10 (about 4k steps)

Wandb: https://wandb.ai/starride-teklia/DAI-CReTDHI/runs/hk78u308

Evaluation

Set	CER (%)	WER (%)	F1 @ 0.0 (%)	F1 @ 0.3 (%)	N samples	N entities
QWEN2.5-VL-7B Flat	10.23	18.07	83.6	91.96	55	808
QWEN2.5-VL-7B Nested	5.48	15.94	84.86	92.27	58	909

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

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Teklia/DAI-cards-nested",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)

processor = AutoProcessor.from_pretrained("Teklia/DAI-cards-nested")

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "12e74aa3-4d7d-47b4-b46b-7013b9a1f251.jpg",
            },
            {"type": "text", "text": "Extrait les informations en XML."},
        ],
    }
]

# 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
)
print(output_text[0])

• Output

<root>
    <Décès>
        <Défunt>
            <Nom>Choisnard</Nom>
            <Prénom>Marie Madelaine</Prénom>
            <Sexe>F</Sexe>
            <DateDeNaissance>23 juillet 1753</DateDeNaissance>
            <LieuDeNaissance>Ambroise (Indre-et-Loire)</LieuDeNaissance>
            <Profession>journalière</Profession>
            <Statut>veuf(ve)</Statut>
        </Défunt>
        <Conjoint>
            <Nom>Rocheriou</Nom>
            <Prénom>Pierre</Prénom>
            <Statut>décédé(e)</Statut>
        </Conjoint>
        <Père>
            <Nom>Choisnard</Nom>
            <Prénom>Michel</Prénom>
        </Père>
        <Mère>
            <Nom>Dubeuf</Nom>
            <Prénom>Louise</Prénom>
        </Mère>
    </Décès>
    <Date>
        <Année>1826</Année>
        <Mois>septembre</Mois>
        <Jour>5</Jour>
    </Date>
</root>

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}
}