earl562/mango-varieties-vlm-v2

MangoAI VLM Training Dataset v2

Dataset: earl562/mango-varieties-vlm-v2 Created: 2025-12-02 License: MIT Base Dataset: earl562/mango-varieties

Overview

This is an enhanced multimodal Vision-Language Model (VLM) training dataset for mango variety identification and analysis. It transforms the original tabular dataset of 3,093 mango images into a comprehensive instruction-following dataset with 9,258 training examples.

Dataset Statistics

• Total Examples: 9,258
• Source Images: 3,093 (average 2.99 examples per image)
• Unique Varieties: 189 mango varieties
• Negative Examples: 445 (4.8%) for hallucination prevention
• Response Length:
  • Mean: 369 characters
  • Median: 377 characters
  • Range: 142-799 characters

Dataset Splits

Split	Examples	Percentage
Train	7,334	79.2%
Validation	842	9.1%
Test	1,082	11.7%

Splits are stratified by variety to ensure representative distribution across all splits. 112 varieties appear in all three splits.

Instruction Types

The dataset includes 9 different instruction types with the following distribution:

Positive Examples (95.2%)

1. Identification (30.2%): Simple variety identification

   • "What variety of mango is this?"
   • Response: 150-250 characters

2. Detailed Description (19.6%): Comprehensive analysis

   • "Provide a detailed description of this mango variety."
   • Response: 400-600 characters

3. Health Assessment (15.0%): Disease and quality evaluation

   • "Is this mango healthy? Are there any signs of disease?"
   • Response: 300-450 characters

4. Comparison (12.4%): Compare with similar varieties

   • "How does this compare to a Kent mango?"
   • Response: 300-500 characters

5. Multi-turn (9.8%): 2-3 turn conversations

   • Sequential questions and answers about the same mango
   • Multiple exchanges building on context

6. Chain-of-Thought (8.1%): Step-by-step reasoning

   • "Walk me through your identification process."
   • Response: 450-600 characters

Negative Examples (4.8%)

7. False Disease Correction (2.0%): Correcting false disease claims

   • User incorrectly identifies disease
   • Model corrects with proper explanation

8. Misidentification Correction (1.9%): Correcting wrong variety ID

   • User incorrectly names variety
   • Model corrects with distinguishing features

9. Uncertainty Acknowledgment (1.0%): Expressing appropriate uncertainty

   • Model acknowledges identification challenges
   • Provides qualified answer with caveats

Message Format

The dataset uses Qwen2-VL compatible message format:

{
  "id": "mango_000001_identification",
  "messages": [
    {
      "role": "system",
      "content": [{"type": "text", "text": "System prompt..."}]
    },
    {
      "role": "user",
      "content": [
        {"type": "image", "image": "https://..."},
        {"type": "text", "text": "What variety of mango is this?"}
      ]
    },
    {
      "role": "assistant",
      "content": [{"type": "text", "text": "This is a Kent mango..."}]
    }
  ],
  "variety": "Kent",
  "task_type": "identification",
  "is_negative": false,
  "source_image_url": "https://...",
  "response_length": 342
}

Variety Knowledge Base

The dataset includes rich domain knowledge for 14+ major varieties:

• Kent, Carrie, Valencia Pride, Haden, Keitt
• Alphonso, Ataulfo, Tommy Atkins, Glenn
• Nam Doc Mai, Pickering, Osteen, Manila, Coconut Cream

Each variety profile includes:

• Origin and history
• Flavor profile (detailed taste description)
• Texture characteristics
• Ripening season
• Fruit size and shape
• Skin coloration
• Key distinguishing features
• Similar varieties for comparison
• Tree characteristics

Quality Features

Rich, Domain-Specific Responses

• All responses include specific variety characteristics (not generic templates)
• Responses incorporate botanical accuracy and horticultural details
• Natural language variation across similar prompts

Hallucination Prevention

• Negative examples teach the model to:
  • Reject false disease claims
  • Correct misidentifications
  • Express appropriate uncertainty
  • Avoid overconfident predictions

Variety Coverage

• All 189 varieties represented across splits
• Balanced representation of common and rare varieties
• Stratified sampling ensures no variety is exclusive to one split

Response Quality

• Target length 300-600 characters (achieved: mean 369)
• No responses under 100 characters
• 227 responses exceed 700 characters for complex queries
• Standard deviation: 164.8 characters (good variation)

Use Cases

This dataset is designed for:

1. Fine-tuning Vision-Language Models (Qwen2-VL, LLaVA, BLIP-2, etc.)
2. Mango variety identification systems
3. Agricultural AI assistants
4. Quality assessment and disease detection
5. Educational applications for fruit identification
6. Multimodal reasoning research

Training Recommendations

Recommended Models

• Qwen2-VL (primary target format)
• LLaVA 1.5/1.6
• BLIP-2
• InstructBLIP
• Any vision-language model supporting image-text interleaved input

Training Configuration

# Recommended hyperparameters
{
  "learning_rate": 2e-5,
  "batch_size": 4,
  "gradient_accumulation_steps": 8,
  "num_epochs": 3,
  "warmup_ratio": 0.1,
  "max_seq_length": 2048,
  "lora_r": 64,
  "lora_alpha": 16,
}

Data Loading

from datasets import load_dataset
import json

# Load dataset
dataset = load_dataset("earl562/mango-varieties-vlm-v2")

# Parse messages
def parse_messages(example):
    example["messages"] = json.loads(example["messages"])
    return example

dataset = dataset.map(parse_messages)

# Use train/validation splits
train_dataset = dataset["train"]
val_dataset = dataset["validation"]

Validation Results

• All example IDs are unique
• All messages have valid format
• Response lengths are appropriate
• No duplicate content detected
• Image URLs are accessible

Limitations

1. Image Quality: Some source images may have varying quality
2. Variety Name Cleaning: Some variety names contain retail suffixes (e.g., "Pre-Order")
3. Knowledge Base Coverage: Only 14 varieties have detailed profiles; others use generic information
4. Synthetic Data: All instruction-response pairs are synthetically generated from source descriptions
5. Single Image per Example: Each example uses one image (no multi-image contexts)

Citation

If you use this dataset, please cite:

@dataset{mangoai_vlm_v2,
  title={MangoAI VLM Training Dataset v2},
  author={Earl Perry},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/earl562/mango-varieties-vlm-v2}
}

Acknowledgments

• Base dataset collected from multiple nursery websites
• Inspired by LLaVA, Qwen-VL, and InstructBLIP training formats
• Built with the HuggingFace datasets library

Updates

v2.0 (2025-12-02):

• Initial release
• 9,258 examples across 9 instruction types
• Stratified splits with 189 varieties
• Qwen2-VL compatible format
• 4.8% negative examples for hallucination prevention

Contact

For questions or issues, please open an issue on the dataset repository or contact via HuggingFace.

---

Dataset Homepage: https://huggingface.co/datasets/earl562/mango-varieties-vlm-v2 Base Dataset: https://huggingface.co/datasets/earl562/mango-varieties License: MIT