CCB/cis5300-language-models

Language Models and City Classification (CIS 5300)

Dataset Description

This dataset supports two tasks for learning about n-gram language models:

1. Shakespeare Text Generation

Raw text from Shakespeare's collected works (~4.5 MB, 167K lines) for training character-level n-gram language models. Students build models of increasing order (n=1 through n=7) and generate Shakespeare-style text.

2. City Name Classification

Classify city names by country of origin using character-level language models. The idea: city names from different countries have distinctive character patterns (e.g., German cities often end in "-burg" or "-stadt", Chinese cities have different character distributions).

Countries: Afghanistan, China, Germany, Finland, France, India, Iran, Pakistan, South Africa

Usage

City Classification Task

from datasets import load_dataset

# Load city name classification data
cities = load_dataset("CCB/cis5300-language-models", "cities")
print(cities)
# DatasetDict({
#     train: Dataset({features: ['city_name', 'country_code', 'country'], num_rows: 12392})
#     validation: Dataset({features: [...], num_rows: 1548})
#     test: Dataset({features: [...], num_rows: 1554})
# })

print(cities["train"][0])
# {'city_name': 'shewah', 'country_code': 'af', 'country': 'Afghanistan'}

Shakespeare Corpus

from huggingface_hub import hf_hub_download

# Download Shakespeare text for language modeling
path = hf_hub_download("CCB/cis5300-language-models", "shakespeare.txt", repo_type="dataset")
with open(path) as f:
    text = f.read()
print(f"Shakespeare corpus: {len(text):,} characters")

Dataset Structure

Cities Config

Split	Examples	Description
train	12,392	City names with country labels for training
validation	1,548	Labeled validation set
test	1,554	Labeled test set for evaluation

Fields:

Field	Type	Description
city_name	string	Name of the city
country_code	string	Two-letter country code (af, cn, de, fi, fr, in, ir, pk, za)
country	ClassLabel	Full country name

Supplementary Files

File	Size	Description
shakespeare.txt	4.5 MB	Complete Shakespeare works for training language models
shakespeare_sonnets.txt	9 KB	Shakespeare's sonnets (additional training data)
nytimes_article.txt	5 KB	News article for cross-domain perplexity evaluation
cities_test_unlabeled.csv	20 KB	Unlabeled test cities (for student predictions)

Source

City name data sourced from GeoNames (CC BY 4.0). Shakespeare text from Project Gutenberg (public domain).

Intended Use

This dataset is used for Homework 3 in CIS 5300: Natural Language Processing at the University of Pennsylvania. Students:

1. Build character-level n-gram language models on Shakespeare
2. Implement smoothing (add-k) and evaluate with perplexity
3. Use interpolation to combine models of different orders
4. Apply language models to classify city names by country of origin

Citation

@misc{cis5300-language-models,
    title = {CIS 5300 Language Models Dataset},
    author = {Callison-Burch, Chris},
    year = {2026},
    publisher = {University of Pennsylvania},
}