Update README.md

README.md

@@ -86,14 +86,14 @@ configs:
 
 ### Dataset Summary
 
-MultiModal Dataset is a curated collection of **112,500 samples** spanning three modalities: text, images, and audio. It combines high-quality web content, image-caption pairs from COCO 2017, and audio samples from AudioSet to enable comprehensive multimodal model training and evaluation.
+MultiModal Dataset is a curated collection of **85,000 samples** spanning three modalities: text, images, and audio. It combines high-quality web content, image-caption pairs from COCO 2017, and audio samples from AudioSet to enable comprehensive multimodal model training and evaluation.
 
 The dataset is organized into three subsets:
 - **fineweb**: 37,500 high-quality web text samples (>8,192 tokens each)
 - **coco**: 37,500 image-caption pairs from COCO 2017 (512x512 resolution)
-- **audioset**: 37,500 audio samples with human-annotated labels
+- **audioset**: 10,000 audio samples with human-annotated labels
 
-All subsets follow a consistent 80/13/7 train/validation/test split for reproducible evaluation across modalities.
+Text and image subsets follow an 80/13/7 train/validation/test split. AudioSet uses 80/15/5 split (8K train, 1.5K valid, 500 test).
 
 ### Supported Tasks
 
@@ -174,18 +174,18 @@ English (en)
 
 ### Data Splits
 
-All three subsets use identical split ratios:
+Text and image subsets use identical 80/13/7 split ratios. AudioSet uses 80/15/5:
 
 | Subset | Train | Validation | Test | Total |
 |--------|-------|------------|------|-------|
 | fineweb | 30,000 | 5,000 | 2,500 | 37,500 |
 | coco | 30,000 | 5,000 | 2,500 | 37,500 |
-| audioset | 30,000 | 5,000 | 2,500 | 37,500 |
-| **TOTAL** | **90,000** | **15,000** | **7,500** | **112,500** |
+| audioset | 8,000 | 1,500 | 500 | 10,000 |
+| **TOTAL** | **68,000** | **11,500** | **5,500** | **85,000** |
 
-**Split methodology**: Samples are deterministically assigned using index-based partitioning:
-- Train: First 80% of samples (indices 0 to 0.8×N)
-- Validation: Next 13% of samples (indices 0.8×N to 0.93×N)
+**Split methodology**: 
+- **fineweb/coco**: Deterministic index-based partitioning (80% train, 13% valid, 7% test)
+- **audioset**: Uses AudioSet's built-in splits (bal_train → train, eval → valid/test)
 - Test: Final 7% of samples (indices 0.93×N to N)
 
 This ensures reproducible splits across different loading sessions.