Handwriting OCR

This Lance-formatted version of the Doctor’s Handwritten Prescription BD dataset contains 4,680 cropped PNG images of handwritten medicine names from Bangladesh. Each row keeps the original image bytes with the medicine and generic-name labels…

View on Hugging Face

Source dataset card and downloadable files for lance-format/handwriting-ocr.

This Lance-formatted version of the Doctor’s Handwritten Prescription BD dataset contains 4,680 cropped PNG images of handwritten medicine names from Bangladesh. Each row keeps the original image bytes with the medicine and generic-name labels, plus deterministic search metadata derived from those labels. The dataset contains three source-preserved splits: train, validation, and test.

[!NOTE] Training note: The same samples appear repeatedly because a model (during training) should learn from multiple different handwritten crops of the same medicine, because they capture different writers, pen strokes, sizes, and image quality. The dataset would need to be shuffled and sampled appropriately before running downstream training tasks

Key features

• Original inline PNG bytes in image, with no image folders or sidecar files required.
• Three preserved splits: 3,120 training, 780 validation, and 780 test examples.
• Medicine annotations: 78 medicine names mapped to 15 generic names.
• Built-in text retrieval over searchable_summary, with scalar indices matching the companion OCR retrieval project.

Splits

Split	Rows	Source directory
train	3,120	Training
validation	780	Validation
test	780	Testing

Schema

Column	Type	Notes
id	string	Stable split-qualified row identifier
image	large_binary	Original inline PNG bytes
medicine_name	string	Source medicine-name label
generic_name	string	Source generic-name label
normalized_text	string	Deterministic normalization of medicine_name
category	string	medication for every source row
is_medical	bool	true for every source row
needs_human_review	bool	false; this conversion does not run OCR confidence assessment
searchable_summary	string	Label-derived text: Medication mention: <medicine> (<generic>)
source_dataset	string	Upstream Kaggle dataset identifier
split	string	Original source split name: Training, Validation, or Testing
image_filename	string	Original PNG filename within the source split

normalized_text, category, is_medical, needs_human_review, and searchable_summary are deterministic conversion metadata. They are not OCR or clinical-model predictions.

Pre-built indices

• INVERTED (FTS) on searchable_summary
• BTREE on id
• BITMAP on category
• BITMAP on needs_human_review

Why Lance?

1. Blazing Fast Random Access: Optimized for fetching scattered rows, making it ideal for random sampling, real-time ML serving, and interactive applications without performance degradation.
2. Native Multimodal Support: Store text, embeddings, and other data types together in a single file. Large binary objects are loaded lazily, and vectors are optimized for fast similarity search.
3. Native Index Support: Lance comes with fast, on-disk, scalable vector and FTS indexes that sit right alongside the dataset on the Hub, so you can share not only your data but also your embeddings and indexes without your users needing to recompute them.
4. Efficient Data Evolution: Add new columns and backfill data without rewriting the entire dataset. This is perfect for evolving ML features, adding new embeddings, or introducing moderation tags over time.
5. Versatile Querying: Supports combining vector similarity search, full-text search, and SQL-style filtering in a single query, accelerated by on-disk indexes.
6. Data Versioning: Every mutation commits a new version; previous versions remain intact on disk. Tags pin a snapshot by name, so retrieval systems and training runs can reproduce against an exact slice of history.

Load with datasets.load_dataset

You can load Lance datasets through the standard Hugging Face datasets interface when you want a streaming sample or your pipeline already uses Dataset / IterableDataset.

import datasets

hf_ds = datasets.load_dataset(
    "lance-format/handwriting-ocr",
    split="train",
    streaming=True,
)
for row in hf_ds.take(3):
    print(row["medicine_name"], row["generic_name"])

Load with LanceDB

LanceDB is the embedded retrieval library built on top of the Lance format (docs), and is the interface most users interact with. It wraps the downloaded dataset as a queryable table with search and filter builders, and is the entry point used by the Search, Curate, Evolve, and Versioning examples below.

import lancedb

db = lancedb.connect("./data")
tbl = db.open_table("train")
print(len(tbl))

Load with Lance

pylance is the Python binding for the Lance format and works directly with lower-level APIs. Use it to inspect the schema, fragments, and pre-built indices.

import lance

ds = lance.dataset("./data/train.lance")
print(ds.count_rows(), ds.schema.names)
print(ds.list_indices())

Tip - for production use, download locally first. Streaming from the Hub works for exploration, but repeated random access and training are faster against a local copy:

hf download lance-format/handwriting-ocr --repo-type dataset --local-dir ./handwriting-ocr

Then run the examples below from ./handwriting-ocr, where ./data contains the downloaded Lance splits.

Search

The bundled full-text index makes medicine and generic-name lookups efficient without an embedding model. searchable_summary combines the two labels, so a text query can return the matching word image and its annotations in one bounded result set.

import lancedb

db = lancedb.connect("./data")
tbl = db.open_table("train")

hits = (
    tbl.search("Paracetamol")
    .select(["id", "medicine_name", "generic_name", "searchable_summary"])
    .limit(10)
    .to_list()
)
for row in hits:
    print(row["medicine_name"], "-", row["generic_name"])

The FTS index operates on label-derived text, not a transcription produced from the image. It is useful for inspecting the handwritten examples associated with known medicine or generic names.

Curate

Filters produce small, explicit candidate sets for experiments. The example selects a bounded set of Paracetamol examples while avoiding image bytes until a later step needs them.

import lancedb

db = lancedb.connect("./data")
tbl = db.open_table("train")

candidates = (
    tbl.search()
    .where("generic_name = 'Paracetamol'")
    .select(["id", "image_filename", "medicine_name", "generic_name"])
    .limit(100)
    .to_list()
)

This list can be reviewed directly, saved as an experiment manifest, or used as the input to an annotation or training workflow.

Evolve

Lance can append new columns without rewriting the original images or annotations. For example, a local copy can add a flag for a specific generic name using a SQL expression, then merge independently produced labels by id.

Note: Mutations require a local copy. The examples in this card assume the dataset was downloaded into the current directory, as shown above.

import lancedb
import pyarrow as pa

db = lancedb.connect("./data")
tbl = db.open_table("train")

tbl.add_columns({"is_paracetamol": "generic_name = 'Paracetamol'"})

review_labels = pa.table({
    "id": ["train_00000"],
    "review_status": ["reviewed"],
})
tbl.merge(review_labels, on="id")

The source columns and pre-built indices remain intact. New labels become available to queries as soon as the write commits.

Train

Projection lets a training loop read only the columns it needs. The image bytes and medicine label below are projected through a standard PyTorch DataLoader; fields added later do not add read cost unless selected.

import lancedb
from lancedb.permutation import Permutation
from torch.utils.data import DataLoader

db = lancedb.connect("./data")
tbl = db.open_table("train")

train_ds = Permutation.identity(tbl).select_columns(["image", "medicine_name"])
loader = DataLoader(train_ds, batch_size=64, shuffle=True, num_workers=4)

for batch in loader:
    # Decode PNG bytes and train an OCR or classification model.
    ...

This keeps the original image representation available for training while avoiding unrelated columns on each epoch.

Versioning

Every Lance mutation creates a version, so the table history can identify the exact split and annotation state used by a retrieval service or experiment. The downloaded dataset can list its history and create tags for reproducible local workflows.

import lancedb

db = lancedb.connect("./data")
tbl = db.open_table("train")

print("Current version:", tbl.version)
print("History:", tbl.list_versions())
print("Tags:", tbl.tags.list())

tbl.tags.create("labels-v1", tbl.version)

A pinned version or tag keeps retrieval and training tied to a known dataset state even as review labels or derived columns are added later.

Materialize a subset

Use a filtered query to materialize a compact local LanceDB table for writes or intensive training. The query streams matching rows and projected columns without materializing the full result set in Python memory.

import lancedb

source_db = lancedb.connect("./data")
source_tbl = source_db.open_table("train")

batches = (
    source_tbl.search()
    .where("generic_name = 'Paracetamol'")
    .select(["id", "image", "medicine_name", "generic_name", "searchable_summary"])
    .to_batches()
)

local_db = lancedb.connect("./paracetamol-subset")
local_db.create_table("train", batches)

The resulting local table can replace ./data in the Evolve, Train, and Versioning examples when a compact writable subset is sufficient.

Citation

Please cite both the upstream dataset and the associated research paper:

@inproceedings{mia2024deep,
  title={A Deep Neural Network Approach with Pioneering Local Dataset to Recognize Doctor's Handwritten Prescription in Bangladesh},
  author={Mia, Abdur Rahim and Chowdhury, Mohammad Abdullah-Al-Sajid and Mamun, Abdullah Al and Ruddra, Aurunave Mollik and Tanny, Nawshin Tabassum},
  booktitle={2024 International Conference on Advances in Computing, Communication, Electrical, and Smart Systems},
  pages={1--6},
  year={2024},
  publisher={IEEE},
  doi={10.1109/iCACCESS61735.2024.10499631}
}

Source dataset: Mamun1113, Doctor’s Handwritten Prescription BD dataset. Associated paper: IEEE Xplore.

License

The upstream Kaggle metadata identifies the database license as Open Database License (ODbL) v1.0 and states that the contents are copyright the original authors. The upstream page also describes the dataset as free for educational and research use. Downstream users should review the upstream dataset page and publication, satisfy the ODbL obligations, and obtain any additional permissions needed for their intended use.