bocchia1122/soccer-iq

Soccer IQ - SFT Dataset

Dataset Description

This dataset contains instruction-response pairs for fine-tuning Large Language Models (LLMs) on soccer coaching, tactics, and training knowledge.

Dataset Summary

• Task: Instruction-following for soccer coaching and tactics
• Language: English
• Format: Instruction-response pairs
• Domain: Soccer coaching, tactical analysis, training drills
• Examples: 4
• Categories: attacking, passing, defending

Intended Use

This dataset is designed for:

• Fine-tuning LLMs for soccer coaching assistance
• Training models to understand tactical concepts
• Building soccer knowledge AI systems
• Instruction tuning for sports-specific applications

Dataset Structure

Data Fields

Each example contains:

• instruction (string): A question, task, or prompt about soccer tactics/coaching
• response (string): Detailed answer or explanation
• category (string): Category of the content (e.g., attacking, defending, passing)
• source (string): Origin of the data

Example

{
  "instruction": "Explain the coaching points for a 4-4-2 pressing drill",
  "response": "Key coaching points include: 1. Angle of approach - ensure the pressing player approaches at an angle that cuts off the most dangerous passing lane. 2. Body shape - maintain a side-on stance to be able to react quickly. 3. Timing - coordinate with teammates to press simultaneously. 4. Recovery run - if the press is beaten, immediate sprint back into defensive shape.",
  "category": "defending",
  "source": "soccer_tactics_pdf"
}

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("bocchia1122/soccer-iq")

# Access examples
print(dataset["train"][0])

Fine-Tuning Example

Using TRL (Transformers Reinforcement Learning)

from datasets import load_dataset
from trl import SFTTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments

# Load dataset
dataset = load_dataset("bocchia1122/soccer-iq")

# Load model
model_name = "mistralai/Mistral-7B-v0.1"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Format function
def format_instruction(example):
    return f"### Instruction:\n{example['instruction']}\n\n### Response:\n{example['response']}"

# Training arguments
training_args = TrainingArguments(
    output_dir="./soccer-iq-model",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    learning_rate=2e-5,
)

# Create trainer
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    formatting_func=format_instruction,
    max_seq_length=2048,
)

# Train
trainer.train()

Using Axolotl

Create config.yml:

base_model: mistralai/Mistral-7B-v0.1

datasets:
  - path: bocchia1122/soccer-iq
    type: alpaca
    
sequence_len: 2048
adapter: lora
lora_r: 16
lora_alpha: 32

num_epochs: 3
learning_rate: 0.00002

Run: axolotl train config.yml

Dataset Creation

This dataset was created by:

1. Extracting text from soccer coaching PDFs
2. Cleaning and segmenting the content
3. Generating instruction-response pairs
4. Manual review and quality control

Source Data

The source material includes professional soccer coaching documents covering:

• Attacking tactics and drills
• Defensive strategies and positioning
• Passing techniques and ball skills
• Training session plans
• Coaching methodologies

Limitations and Biases

• Dataset size is limited; may benefit from additional examples
• Content is primarily focused on European football coaching methods
• May not cover all tactical systems or coaching philosophies comprehensively
• Quality of instruction-response pairs depends on source material clarity

Citation

If you use this dataset in your work, please cite:

@dataset{soccer_iq_sft,
  title={Soccer IQ - SFT Dataset},
  author={Dataset Creator},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/bocchia1122/soccer-iq}
}

License

MIT License - See LICENSE file for details

Contact

For questions or issues with this dataset, please open an issue on the dataset repository page.

---

Note: This is a specialized dataset for soccer coaching. For best results, fine-tune on a base model with good general reasoning capabilities.