subhrokomol/siglip2-large-lora-v1-dataset

SigLIP2-Large LoRA v1 — Training Pairs

Material/surface visual-similarity training pairs for fine-tuning SigLIP2-large with LoRA + Supervised Contrastive loss. Each row links a catalog product image to one polygon-clipped material crop from a SAM3-segmented interior room render.

Statistics

Split	Rows
train	5,618
eval	576

Layout

.
├── dataset.jsonl       train split (one JSON object per line)
├── eval.jsonl          held-out eval split
└── segments.tar.gz     image archive — extract to a folder named "segments"
                        in the same dir as dataset.jsonl

Row schema

{
  "image_path": "segments/<room_id>_<seg_id>.jpg",
  "group_id":   "<product_id>",
  "domain":     "<slot_label>",
  "product_id": "<product_id>",
  "room_id":    "<id>",
  "sam3_score": 0.94,
  "match_score": 1.09
}

• image_path — relative path to the cropped material image (after extracting segments.tar.gz)
• group_id — class label for Supervised Contrastive loss; same product = positive pair
• domain — slot category (wall, floors, sofa, etc.)
• sam3_score — SAM3 segmentation confidence
• match_score — Voyage-AI multimodal similarity used during mining

How the data was produced

Pairs were mined by running SAM3 segmentation on interior room renders, generating multimodal embeddings of each segment via voyage-multimodal-3.5, performing ANN search against a catalog of product images, and keeping top-k matches that passed a similarity threshold.

Companion model

subhrokomol/siglip2-large-lora-v1 — LoRA-adapted SigLIP2 trained on this dataset.

Usage

from datasets import load_dataset

ds = load_dataset("subhrokomol/siglip2-large-lora-v1-dataset")
print(ds)
# DatasetDict({'train': Dataset(...), 'eval': Dataset(...)})

To use the image files, extract the tarball:

huggingface-cli download subhrokomol/siglip2-large-lora-v1-dataset \
  segments.tar.gz --repo-type dataset --local-dir .
tar xzf segments.tar.gz

License

Apache 2.0.