metadata

license: mit
task_categories:
  - image-to-text
language:
  - en
tags:
  - handwritten-digits
  - math-education
  - ocr
  - optical-character-recognition
  - handwriting-recognition
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: session_id
      dtype: string
    - name: question_id
      dtype: string
    - name: timestamp
      dtype: string
    - name: operand_a
      dtype: int64
    - name: operand_b
      dtype: int64
    - name: operation
      dtype: string
    - name: correct_answer
      dtype: int64
    - name: difficulty
      dtype: string
    - name: ocr_prediction
      dtype: string
    - name: ocr_parsed_number
      dtype: int64
    - name: is_correct
      dtype: bool
    - name: ocr_model_name
      dtype: string
    - name: ocr_processing_time
      dtype: float64
    - name: ocr_confidence
      dtype: float64
    - name: session_duration
      dtype: int64
    - name: session_total_questions
      dtype: int64
    - name: app_version
      dtype: string
    - name: hardware
      dtype: string
    - name: handwriting_image
      dtype: image
    - name: session_accuracy
      dtype: float64
    - name: session_total_ocr_time
      dtype: float64
    - name: session_avg_ocr_time
      dtype: float64
  splits:
    - name: train
      num_bytes: 4394274.25
      num_examples: 1414
  download_size: 4239274
  dataset_size: 4394274.25

CalcTrainer Dataset 🧮

Handwritten mathematical answers collected from the CalcTrainer interactive math training application.

Dataset Fields

Core Data

Field	Type	Description
`handwriting_image`	Image	Handwritten answer image (~100x100px)
`ocr_prediction`	string	Raw OCR output text
`ocr_parsed_number`	int32	Cleaned numeric value from OCR
`is_correct`	bool	Whether OCR matches correct answer

Mathematical Context

Field	Type	Description
`operand_a`	int32	First number (e.g., 7 in "7 × 3")
`operand_b`	int32	Second number (e.g., 3 in "7 × 3")
`operation`	string	Operation: `+`, `-`, `×`, `÷`
`correct_answer`	int32	Expected correct answer
`difficulty`	string	`Facile` (Easy) or `Difficile` (Hard)

OCR Metrics

Field	Type	Description
`ocr_model_name`	string	OCR model used (e.g., "microsoft/trocr-base-handwritten")
`ocr_processing_time`	float32	Processing time in seconds
`hardware`	string	Hardware used for OCR

Session Info

Field	Type	Description
`session_id`	string	Unique session identifier
`question_id`	string	Unique question identifier
`timestamp`	string	When the session was completed
`session_duration`	int32	Session length (30 or 60 seconds)
`session_accuracy`	float32	Overall session accuracy percentage
`session_avg_ocr_time`	float32	Average OCR time per image in session

Usage

from datasets import load_dataset

dataset = load_dataset("hoololi/CalcTrainer_dataset")
train_data = dataset["train"]

# Example: Access first item
item = train_data[0]
print(f"Math problem: {item['operand_a']} {item['operation']} {item['operand_b']} = {item['correct_answer']}")
print(f"OCR predicted: '{item['ocr_prediction']}' → {item['ocr_parsed_number']}")
print(f"Correct: {item['is_correct']}")

Data Source

Real handwriting samples from users solving math problems in the CalcTrainer application. Users write answers on a digital canvas during timed math sessions.

Generated from: CalcTrainer Interactive Math Training 🧮