license: cc-by-4.0
tags:
- dataset
- document-ocr
- multimodal3
- pipeline
- metadata-sample
Document OCR Multimodal3 Data Notes
Overview
A documented Document OCR data-preparation workflow for Multimodal3 records. The bundled rows demonstrate the schema and validation path rather than pretending to be a full training corpus.
Data sources
- external sources are referenced, not redistributed
- record retrieval date and filtering rules
- keep provenance explicit before any training run
Environment
| Item | Value |
|---|---|
| Python | >= 3.9 |
| torch | 2.x |
| safetensors | any recent |
Validation checklist
Before using the prepared data, verify source licenses, duplicates across splits, missing values, label balance, and modality-specific corruption. Record the source version and every filtering rule so a later run can reproduce the same rows.
Risks and limitations
External sources may be unrepresentative, mis-licensed, or contain sensitive fields. Source review, privacy checks and bias analysis are the user's responsibility before training or redistribution.
Processing choices
| Item | Value |
|---|---|
| Storage format | jsonl |
| Preprocessing | curriculum |
| Augmentation | heavy |
| Split strategy | leave one out |
| Sampling | active |
| Quality checks | adaptive |
| Labeling | manual |
Included material
dataloader.py— loading, cleaning and split preparation.dataset_infos.json— schema and split metadata.metadata_sample.jsonl— readable records for schema checks.
Intended use
The repository is suitable for testing the data pipeline, adapting it to a documented source, and preparing controlled research splits. Release status: metadata sample; full source data not bundled. The sample is for schema inspection only and should not be reported as a full training corpus.
Files
| File | Purpose |
|---|---|
dataloader.py |
primary artifact |
README.md |
this documentation |
dataset_infos.json |
schema metadata |
metadata_sample.jsonl |
schema sample |
License
Distributed under the cc-by-4.0 license. External datasets keep their own terms — check them separately.