dddraxxx/filtered_deepseek_v31_referring_expression_parsing

DeepSeek v3.1 Quality-Filtered Referring Expression Parsing + Distractor Labels

This dataset contains parsed referring expressions from the RefCOCO, RefCOCOg, and RefCOCO+ validation sets, processed using DeepSeek v3.1 with quality filtering, plus corresponding distractor label annotations in COCO format.

Dataset Description

Overview

• Model: DeepSeek v3.1 (deepseek-chat)
• Processing: Quality-filtered results from referring expression parsing
• Datasets: RefCOCO, RefCOCOg, RefCOCO+ validation splits
• Total Records: 21,955 parsed referring expressions + distractor annotations
• Additional: COCO-format distractor label files for visual grounding research

Parsed Expression Files

1. deepseek_v31_refcoco_quality_filtered.jsonl (9,241 records, 3.9MB)
   • Mixed RefCOCO validation data with parsed noun phrases
2. deepseek_v31_refcocog_quality_filtered.jsonl (4,668 records, 2.5MB)
   • RefCOCOg validation subset with hierarchical noun parsing
3. deepseek_v31_refcocoplus_quality_filtered.jsonl (8,046 records, 3.6MB)
   • RefCOCO+ validation subset with dependency-level parsing

Distractor Label Files (COCO Format)

4. refcoco_val_with_distractor_labels_coco.json (18MB, 10,834 annotations)
   • RefCOCO validation with distractor object annotations
5. refcocog_val_with_distractor_labels_coco.json (8.0MB, ~4,668 annotations)
   • RefCOCOg validation with distractor object annotations
6. refcocoplus_val_with_distractor_labels_coco.json (18MB, ~8,046 annotations)
   • RefCOCO+ validation with distractor object annotations

Data Structure

Parsed Expression Format

Each JSONL record contains:

{
  "dataset": "refcoco|refcocog|refcocoplus",
  "split": "validation", 
  "image_id": 580957,
  "ref_id": 77,
  "ann_id": 1537681,
  "sent_id": 222,
  "image_path": "coco/train2014/COCO_train2014_000000580957.jpg",
  "target_bbox": [468.3, 0.91, 640.0, 117.03],
  "sent": "bowl behind the others can only see part",
  "tokens": ["bowl", "behind", "the", "others", "can", "only", "see", "part"],
  "ref_length": 8,
  "parsed_nouns": [
    {"text": "bowl", "start": 0, "end": 1, "level": 0},
    {"text": "others", "start": 3, "end": 4, "level": 1}
  ],
  "target_indices": [0]
}

Distractor Label Format (COCO JSON)

Standard COCO format with additional distractor annotations:

{
  "info": {"description": "COCO 2014 Dataset", "version": "1.0", ...},
  "licenses": [...],
  "images": [{"id": 580957, "file_name": "COCO_train2014_000000580957.jpg", ...}],
  "annotations": [
    {
      "id": 1537681,
      "image_id": 580957,
      "category_id": 51,
      "bbox": [468.3, 0.91, 171.7, 116.12],
      "area": 19942.604,
      "iscrowd": 0,
      "is_distractor": false
    },
    {
      "id": 1537682,
      "image_id": 580957, 
      "category_id": 51,
      "bbox": [151.96, 139.46, 454.93, 283.73],
      "area": 129063.7,
      "iscrowd": 0,
      "is_distractor": true
    }
  ],
  "categories": [...]
}

Key Fields

• parsed_nouns: Hierarchically structured noun phrases with dependency levels
• level: Dependency level (0=target, 1=primary modifier, 2=secondary, etc.)
• target_indices: Indices pointing to target noun phrases in parsed_nouns array
• is_distractor: Boolean flag indicating if object is a distractor (true) or target (false)

Quality Filtering

The results have been quality-filtered to ensure:

• Accurate noun phrase extraction
• Proper hierarchical dependency parsing
• Valid token span alignment
• Meaningful target identification
• Proper distractor vs target object labeling

Usage

Loading Parsed Expressions

import json

# Load the dataset
with open('deepseek_v31_refcoco_quality_filtered.jsonl', 'r') as f:
    data = [json.loads(line) for line in f]

# Example: Extract target nouns
for record in data:
    target_nouns = [record['parsed_nouns'][i]['text'] 
                   for i in record['target_indices']]
    print(f"Sentence: {record['sent']}")
    print(f"Target nouns: {target_nouns}")

Loading Distractor Labels

import json

# Load COCO-format distractor labels
with open('refcoco_val_with_distractor_labels_coco.json', 'r') as f:
    coco_data = json.load(f)

# Filter target vs distractor objects
targets = [ann for ann in coco_data['annotations'] if not ann.get('is_distractor', False)]
distractors = [ann for ann in coco_data['annotations'] if ann.get('is_distractor', False)]

print(f"Target objects: {len(targets)}")
print(f"Distractor objects: {len(distractors)}")

Research Applications

This dataset is valuable for:

• Referring Expression Comprehension: Training models to understand natural language descriptions
• Visual Grounding: Learning to ground language in visual scenes
• Distractor Analysis: Studying how models handle similar objects in complex scenes
• Noun Phrase Parsing: Understanding hierarchical language structure
• Multi-modal Learning: Combining vision and language understanding

Citation

If you use this dataset, please cite the original RefCOCO datasets and acknowledge the parsing methodology:

@misc{deepseek_v31_referring_parsing_2025,
  title={DeepSeek v3.1 Quality-Filtered Referring Expression Parsing with Distractor Labels},
  author={dddraxxx},
  year={2025},
  howpublished={\url{https://huggingface.co/datasets/dddraxxx/filtered_deepseek_v31_referring_expression_parsing}}
}

License

MIT License - See original RefCOCO dataset licenses for underlying data usage terms.