ActivityNet-FG-It

Dataset Description

ActivityNet-FG-It (ActivityNet Factual Grounding for Instruction Tuning) is a synthetic dataset designed for training video caption evaluation models. It contains 44K video-caption pairs with systematically controlled factual errors, graded quality scores (1-5), and natural language explanations identifying the errors.

This dataset was created to address the lack of captions with varying degrees of factual quality for training reference-free video caption evaluators.

Dataset Summary

Attribute	Value
Total Samples	44,000
Samples per Score	8,800 (balanced)
Source Dataset	ActivityNet Captions (train split)
Score Range	1-5 (discrete)
Generator LLM	Llama-3.3-70B-Instruct

Dataset Creation

Generation Pipeline

The dataset is generated through a systematic pipeline that creates captions with controllable factual errors:

Ground Truth Caption
        ↓
   [LLM Parser]
        ↓
Objects: [obj1, obj2, ...]    Actions: [act1, act2, ...]
        ↓                              ↓
   [LLM Generator]               [LLM Generator]
        ↓                              ↓
Alternative Objects           Alternative Actions
        ↓                              ↓
        └──────────┬───────────────────┘
                   ↓
            [LLM Substitution]
                   ↓
            Pseudo Caption
                   ↓
            Quality Score + Explanation

Caption Generation Process

1. Object Extraction: Parse objects from ground-truth captions (e.g., "man", "guitar", "field")
2. Action Extraction: Parse actions from ground-truth captions (e.g., "playing", "standing")
3. Alternative Generation:
   • For objects: Generate semantically different alternatives from the same category (e.g., "guitar" → "violin")
   • For actions: Generate different actions the subject could perform (e.g., "playing" → "holding")
4. Random Substitution: Randomly select K objects and L actions to replace
5. Caption Synthesis: Use LLM to substitute selected elements while maintaining grammatical correctness

Scoring Mechanism

Quality scores are computed deterministically based on factual accuracy:

score = 1 - (# of changed objects & actions) / (total # of objects & actions)

The continuous score is then discretized to a 1-5 scale:

discrete_score = round(score × 4 + 1)

Data Format

Each sample contains:

{
  "id": "v_QOlSCBRmfWY_0_1",
  "messages": [
    {
      "role": "user",
      "content": "<image><image>...<image>\n<caption>A woman is releasing a hula hoop and talking.</caption>\n\nYou are given a video and a caption describing the video content. Please rate the helpfulness, relevance, accuracy, level of details of the caption. The overall score should be on a scale of 1 to 5, where a higher score indicates better overall performance. Please first output a single line containing only one integer indicating the score. In the subsequent line, please provide a comprehensive explanation of your evaluation, avoiding any potential bias. STRICTLY FOLLOW THE FORMAT."
    },
    {
      "role": "assistant",
      "content": "2\nThe caption does not accurately capture the video content. For example, the objects (woman, hula hoop) and actions (releasing) are incorrect."
    }
  ],
  "images": [
    "/path/to/frame_001.jpg",
    "/path/to/frame_002.jpg",
    "..."
  ]
}

Field Descriptions

Field	Type	Description
id	string	Unique identifier (video_id + segment_id + sample_id)
messages	list	Conversation format with user query and assistant response
images	list	Paths to video frames (32 frames uniformly sampled)

Message Format

User Message:

• <image> tokens (repeated for each frame)
• <caption> tags containing the candidate caption
• Evaluation instruction prompt

Assistant Message:

• Line 1: Quality score (integer 1-5)
• Line 2+: Natural language explanation

Explanation Templates

Explanations follow structured templates based on error types:

Error Type	Template
No errors	"The caption is helpful, relevant, accurate, and informative to the video content."
Objects only	"The caption does not accurately capture the video content. For example, the objects ({wrong_objects}) are incorrect."
Actions only	"The caption does not accurately capture the video content. For example, the actions ({wrong_actions}) are incorrect."
Both	"The caption does not accurately capture the video content. For example, the objects ({wrong_objects}) and actions ({wrong_actions}) are incorrect."

Dataset Statistics

Score Distribution (Balanced)

Score	Count	Percentage
1	8,800	20%
2	8,800	20%
3	8,800	20%
4	8,800	20%
5	8,800	20%
Total	44,000	100%

Augmentation Types

Type	Description
change_obj	Only objects are substituted
change_act	Only actions are substituted
change_obj_act	Both objects and actions are substituted (primary)

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("dipta007/ActivityNet-FG-It")

# Access train split
train_data = dataset["train"]

# Example sample
sample = train_data[0]
print(f"ID: {sample['id']}")
print(f"Score: {sample['messages'][1]['content'].split(chr(10))[0]}")
print(f"Num frames: {len(sample['images'])}")

Citation

If you use this dataset, please cite our paper:

@misc{dipta2025advancingreferencefreeevaluationvideo,
      title={Advancing Reference-free Evaluation of Video Captions with Factual Analysis},
      author={Shubhashis Roy Dipta and Tz-Ying Wu and Subarna Tripathi},
      year={2025},
      eprint={2509.16538},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.16538},
}

Acknowledgements

• ActivityNet Captions for the source video-caption data
• Meta AI for the Llama model used in caption generation