NealCaren/inkbench-4b-baseline-detailed

InkBench Transcriptions

This dataset contains machine-generated transcriptions of the InkBench dataset.

Model Used

• Model: unsloth/Qwen3-VL-4B-Instruct-unsloth-bnb-4bit
• Source Dataset: NealCaren/InkBench
• Number of Examples: 400

Dataset Structure

Each example contains:

• image_name: Original image filename
• type: Document type
• text: Ground truth transcription (from InkBench)
• transcription: Model-generated transcription
• Any other metadata fields from the original dataset

The original images are NOT included to reduce dataset size. Use this dataset with the original InkBench dataset for evaluation.

Usage

from datasets import load_dataset

# Load transcriptions
ds = load_dataset("NealCaren/inkbench-4b-baseline-detailed")

# Compare ground truth vs prediction
for example in ds["train"]:
    print(f"Ground truth: {example['text']}")
    print(f"Predicted: {example['transcription']}")

Evaluation

To evaluate these transcriptions:

from Levenshtein import distance as levenshtein_distance

def calculate_cer(pred, gt):
    return levenshtein_distance(pred, gt) / len(gt) if len(gt) > 0 else 0

# Calculate CER for each example
for example in ds["train"]:
    cer = calculate_cer(example["transcription"], example["text"])
    print(f"{example['image_name']}: CER = {cer:.2%}")

Citation

If you use this dataset, please cite both InkBench and the model used for transcription.