Krishna0812's picture
Updating Readme to explain dataset , application
ddaf1f2 verified
|
Raw
History Blame Contribute Delete
3.51 kB
metadata
license: mit
task_categories:
  - text-generation
  - translation
language:
  - en
  - hi
tags:
  - stories
  - parallel-corpus
  - hindi
  - english
  - simple-stories
  - synthetic
size_categories:
  - 1M<n<10M

Simple-Stories-Eng-Hin

A parallel English–Hindi dataset of short, simple stories. Each row contains the same story written in English and its Hindi counterpart, making it useful for training or evaluating small language models, translation systems, and Hindi text generation models on simple narrative text.

Dataset Details

  • Rows: ~1.72M
  • Format: JSON (auto-converted to Parquet)
  • Split: train (single split)
  • Size: ~9.38 GB
  • License: MIT

Columns

Column Type Description
english string A short story in English (~231–2,570 characters)
hindi string The corresponding story in Hindi (~186–3,040 characters)

Dataset Structure

Each example is a story pair covering everyday, fairy-tale, and adventure-style narratives — talking animals, curious children, magical objects, treasure maps, and similar simple, self-contained plots. Stories are short (roughly a paragraph to a few paragraphs) and written in easy-to-follow language, similar in spirit to the TinyStories / SimpleStories style of datasets.

Example (truncated):

{
  "english": "Key turned in the old lock. A girl named Lily held her breath...",
  "hindi": "चाबी पुराने ताले में घूम गई। लिली नाम की एक लड़की ने अपनी सांस रोकी..."
}

Applications & Use Cases

Language modeling

  • Pretraining or fine-tuning small/efficient LLMs (SimpleStories/TinyStories-style) for Hindi or bilingual generation
  • Studying how small models acquire grammar and narrative coherence in a low-resource language like Hindi
  • Distillation: using story pairs as a compact training signal for smaller student models

Machine translation

  • Training or fine-tuning English↔Hindi translation models
  • Evaluating translation quality on simple, unambiguous sentence structures where errors are easy to spot
  • Building baselines before scaling to harder, domain-specific parallel corpora

Evaluation & benchmarking

  • Benchmarking Hindi text generation fluency and coherence
  • Testing cross-lingual transfer (e.g. English-pretrained model fine-tuned on Hindi)
  • Sanity-checking tokenizer or vocabulary coverage for Devanagari script

Education & accessibility

  • Bilingual reading practice or language-learning apps (parallel sentence/story display)
  • Generating graded reading material for Hindi learners
  • Text-to-speech or read-along tools needing simple, clean narrative text

Other

  • Data augmentation for children's-story or creative-writing generators in Hindi
  • Seed data for synthetic dataset generation (e.g. prompting larger models to write more stories in this style)

Notes

  • The Hindi text is a translation of the English story rather than an independently authored story, so both columns describe the same narrative.
  • Given the simple, formulaic story style, users working with this data for translation quality benchmarks should be aware the vocabulary and sentence structure are intentionally simple.

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("bhasha-labs/Simple-Stories-Eng-Hin")
print(ds["train"][0])

License

Released under the MIT License.