Fine-tuning for OCR task

[MohamedLahmeri01]

Search before asking

• I have searched the Multimodal Maestro issues and found no similar feature requests.

Question

can it be used to fine-tune Qwen-2.5 VL for ocr tasks ? if yes , is their a guide to use ?
provide the dataset structure for fine-tuning .

Additional

No response

[onuralpszr]

@MohamedLahmeri01 yes and google collab codes are already in the repo as well, here is links;

https://colab.research.google.com/github/roboflow/maestro/blob/develop/cookbooks/maestro_qwen2_5_vl_json_extraction.ipynb

https://github.com/roboflow/maestro/blob/develop/cookbooks/maestro_qwen2_5_vl_json_extraction.ipynb

I hope that helps

[onuralpszr]

Converting to Q&A discussion

[MohamedLahmeri01]

i have built a custom dataset that i wanted to try on it , is their any specific dataset structure should i follow ?

[MohamedLahmeri01]

this is my dataset structure : [
{
"query": "\nOCR with format {AIDA} :",
"response": "طفنا وسعينا مع شيخ. كان جاري في الخيمة يتكلم وهو نائم بكلمات لا أفهمها",
"images": [
"im10418.jpg"
]
},
{
"query": "\nOCR with format {AIDA} :",
"response": "مثل انقض بغـلس له الضابط لزمتك. سألت راجح هل بلغ أصحابنا",
"images": [
"im10419.jpg"
]
}
]

[onuralpszr]

You can follow "JSONL format" plus the links I posted also uses JSONL as well.

[MohamedLahmeri01]

@onuralpszr , i will give it a try and then returns to you.

[SkalskiP]

Hi @MohamedLahmeri01 👋🏻 I added a docs page this morning diving deeper into the JSONL format. I hope it will be helpful.