README.md

CS701 Course Project Dataset

For task "ViT Model Adaptation"

Overview

This is an anonymized image classification dataset containing 10,000 classes.

Dataset Statistics

• Number of classes: 10,000
• Training samples: 450,000 (45 per class, with labels)
• Validation samples: 50,000 (5 per class, without labels)
• Test samples: 100,000 (10 per class, without labels)

Dataset Structure

train/
  class_0000/  # 45 images
  class_0001/  # 45 images
  ...
  class_9999/  # 45 images

val/
  image_000000.jpg
  image_000001.jpg
  ...
  image_049999.jpg

test/
  image_000000.jpg
  image_000001.jpg
  ...
  image_099999.jpg

Files

• train.txt: Training image paths and labels (format: train/class_XXXX/filename.jpg class_id)
• val.txt: Validation image paths only (format: val/image_XXXXXX.jpg)
• test.txt: Test image paths only (format: test/image_XXXXXX.jpg)
• metadata.json: Dataset metadata

Benchmark Submission

1. Train your model on the training set (450,000 samples with labels)
2. Generate predictions on validation set (50,000 samples without labels)
3. Generate predictions on test set (100,000 samples without labels)
4. Submit predictions to CodaBench for evaluation.

Prediction Format

Your prediction files should follow this format:

val_predictions.txt:

val/image_000000.jpg 1234
val/image_000001.jpg 5678
...
val/image_049999.jpg 9012

test_predictions.txt:

test/image_000000.jpg 1234
test/image_000001.jpg 5678
...
test/image_099999.jpg 9012

Each line contains: <image_path> <predicted_class_id> where class_id is in range [0, 9999].

Evaluation

Submit your prediction files to the dataset maintainer for evaluation. The evaluation will compute:

• Overall accuracy
• Per-class accuracy
• Additional metrics (if applicable)

Note: Validation and test labels are withheld to prevent overfitting and ensure fair benchmarking.