technicalheist/cricket-alpaca

Cricket Match Alpaca Dataset

This dataset contains cricket match information formatted for instruction-tuning of Large Language Models (LLM) in Alpaca format.

Dataset Splits

Split	Matches	Entries	Percentage
Train	44,616	5,353,920	80%
Valid	5,577	669,240	10%
Test	5,578	669,360	10%

• Split Method: Match-level split (all 120 questions for a match go to the same split)
• Random Seed: 42
• No Data Leakage: Matches are not shared across splits

Data Format

Each entry is a JSON object with the following structure:

{
  "instruction": "Question text with match-specific details",
  "input": "Context: Tournament: X, Teams: Y vs Z, Scores: A-B...",
  "output": "Answer based on the match data"
}

Example Entry

{
  "instruction": "What is the outcome of the cricket match between VB Kanchi Veerans and Karaikudi Kaalai on 2021-07-27?",
  "input": "Tournament: Tamil Nadu Premier League, Regular Season, Home: VB Kanchi Veerans (Kanchi Veerans), Away: Karaikudi Kaalai (Karaikudi Kaalai), Scores: 148-149, Country: India (IN), Winner: Karaikudi Kaalai, Note: ",
  "output": "Karaikudi Kaalai won against VB Kanchi Veerans by 1 runs on 2021-07-27. Scores: Karaikudi Kaalai 149, VB Kanchi Veerans 148."
}

Question Categories

1. Basic Outcome (20 questions)

Questions about match results, winners, and outcomes.

2. Team Information (20 questions)

Questions about team names, short names, codes, and slugs.

3. Tournament Details (20 questions)

Questions about tournament names, slugs, and categories.

4. Geographic (20 questions)

Questions about country names, alpha2, and alpha3 codes.

5. Match Metadata (20 questions)

Questions about match dates, status, seasons, and notes.

6. Scores and Margin (20 questions)

Questions about scores, margins, and run differences.

Fields Available

Field	Description	Example
match_date	Date of the match	2021-07-27
home_team_name	Home team full name	VB Kanchi Veerans
away_team_name	Away team full name	Karaikudi Kaalai
home_shortName	Home team short name	Kanchi Veerans
away_shortName	Away team short name	Karaikudi Kaalai
home_nameCode	Home team code	KAN
away_nameCode	Away team code	KKA
home_slug	Home team slug	vb-kanchi-veerans
away_slug	Away team slug	karaikudi-kaalai
home_score	Home team score	148
away_score	Away team score	149
winner_code	Winner (1=home, 2=away, 3=draw)	2
tournament_name	Tournament name	Tamil Nadu Premier League, Regular Season
tournament_slug	Tournament slug	tamil-nadu-premier-league
category_name	Tournament category	India
country_name	Country name	India
country_alpha2	Country code (2-letter)	IN
country_alpha3	Country code (3-letter)	IND
season_name	Season name	2021
note	Match notes	(filtered for relevance)
status_description	Match status	Ended

Data Source

• Source: Trusted cricket data provider (downloaded dataset)
• Database: SQLite (dataset/cricket.db)
• Tables Used:
  • matches - Core match data
  • teams - Team information (names, slugs, shortNames, nameCodes)
  • countries - Country information (names, alpha2, alpha3)
  • tournaments - Tournament details (names, slugs)
  • categories - Tournament categories

Data Quality

Quality Checks Passed

• ✅ All entries are valid JSON
• ✅ No empty instructions, inputs, or outputs
• ✅ No unreplaced placeholders
• ✅ Integer margins (not floats like "1.0")

Known Limitations

• Missing Country Data: 28% of matches (15,682) lack country information (country_name, alpha2, alpha3)
• Missing Scores: 7% of matches lack home_score or away_score
• Note Field: Notes are filtered to only include relevant ones (mentioning teams in the match). Original notes may not match the actual match teams in the source database.

Usage

Loading with Datasets Library

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/cricket-alpaca")

# Access splits
train_data = dataset['train']
valid_data = dataset['validation']
test_data = dataset['test']

print(f"Train: {len(train_data)} examples")
print(f"Valid: {len(valid_data)} examples")
print(f"Test: {len(test_data)} examples")

# Access an example
example = train_data[0]
print(f"Instruction: {example['instruction']}")
print(f"Input: {example['input']}")
print(f"Output: {example['output']}")

Loading from JSONL Files

import json

def load_jsonl(file_path):
    data = []
    with open(file_path, 'r') as f:
        for line in f:
            data.append(json.loads(line))
    return data

# Load training data
train_data = load_jsonl('train.jsonl')
print(f"Loaded {len(train_data)} training examples")

Fine-tuning Format

The dataset is already in Alpaca format, compatible with:

• LLaMA Fine-tuning
• Alpaca-style instruction tuning
• Most instruction-tuning frameworks

Repository Structure

.
├── train.jsonl          # Training split (5.35M entries)
├── valid.jsonl          # Validation split (669K entries)
├── test.jsonl           # Test split (669K entries)
├── README.md            # This file (dataset card)
└── dataset/
    └── cricket.db      # SQLite database with raw data

Scripts

The dataset was generated using the following scripts:

Script	Description
generate_split_dataset.py	Generates train/val/test splits from database
generate_dataset.py	Generates single dataset file
export_clean_csv.py	Exports clean CSV from database
quality_check.py	Runs quality checks on generated data

Reproducibility

• Python Version: 3.x
• Dependencies: sqlite3, json, random, datasets, huggingface-hub
• Random Seed: 42 (for reproducible splits)
• Database: dataset/cricket.db (provided)

Citation

If you use this dataset, please cite:

@dataset{cricket_alpaca_2024,
  author = {Your Name},
  title = {Cricket Match Alpaca Dataset},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/your-username/cricket-alpaca}}
}

License

MIT License

Contact

For questions or issues, please open an issue in the repository.