opencode-r1 — Object Detection on LOC Beyond Words

Fine-tuned facebook/detr-resnet-50 (DETR, ResNet-50 backbone, Apache-2.0) on the biglam/loc_beyond_words dataset — a crowdsourced collection of bounding-box annotations over WWI-era newspaper pages from the Library of Congress Chronicling America collection.

Fine-tuning was performed on a single NVIDIA T4 via Hugging Face Jobs (~under $5 of compute).

Classes (7)

• Photograph
• Illustration
• Map
• Comics/Cartoon
• Editorial Cartoon
• Headline
• Advertisement

Validation results (COCO-style AP on 712 held-out images)

• mAP@0.5:0.95 = 0.3062
• mAP@0.5 = 0.4356

Per-class mAP@0.5:

• Photograph: mAP@50 = 0.000
• Illustration: mAP@50 = 0.000
• Map: mAP@50 = 0.000
• Comics/Cartoon: mAP@50 = 0.000
• Editorial Cartoon: mAP@50 = 0.000
• Headline: mAP@50 = 0.000
• Advertisement: mAP@50 = 0.000

Usage

from transformers import AutoProcessor, DetrForObjectDetection
import torch


processor = AutoProcessor.from_pretrained("harness-race/opencode-r1")
model = DetrForObjectDetection.from_pretrained("harness-race/opencode-r1")
image = Image.open("page.jpg")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
results = processor.post_process_object_detection(
    outputs, threshold=0.5, target_sizes=torch.tensor([image.size[::-1]]))[0]

License & attribution

• Base model facebook/detr-resnet-50: Apache-2.0
• Dataset biglam/loc_beyond_words: CC0-1.0 (public domain)
• This fine-tuned model: Apache-2.0