yonaitay/Scouter-ai-dataset

Scouter AI Dataset

This dataset contains soccer scouting reports and player metrics.

Dataset Schema

Column	Type
Name	string
Age	int64
Position	string
Team	string
League	string
Nation	string
Rating	int64
Market_Value	string
Scout_Report	string

⚽ AI-Powered Synthetic Soccer Scout Dataset

Overview

This project generates a high-quality synthetic dataset of 20,000 professional soccer players using Python, Faker, and a Large Language Model (LLM).

Note: Per my professor's recommendation, this project utilizes Qwen-2.5-0.5B-Instruct for text generation. This model was selected for its efficiency and ability to generate unique, context-aware textual scouting reports for every player, making it ideal for Natural Language Processing (NLP) tasks, regression analysis, and data visualization.

🚀 Features

• Massive Scale: Generates 20,000 unique player profiles.
• Realistic Attributes: Includes Name, Age, Position, Team, League, and Nationality.
• Smart Valuation Logic: Custom algorithm calculates Market Value based on Age, Rating, and potential.
• Tiered Skill Distribution: Simulates real-world talent distribution (Tier 1 nations have higher probability of elite players).
• AI-Generated Text: Uses Qwen/Qwen2.5-0.5B-Instruct to write professional one-sentence scouting reports for every player.

🛠️ Technology Stack

• Language: Python
• Environment: Google Colab (Requires T4 GPU)
• Libraries: transformers, torch, pandas, faker, accelerate, adjustText

📊 Dataset Structure

The output file (soccer_scout_data.csv) contains the following columns:

Column	Description
Name	Synthetic player name
Age	Integer between 17 and 38
Position	E.g., Striker, Center Back, GK
Team	Assigned from real-world leagues (Premier League, La Liga, etc.)
League	The league the team belongs to
Nation	Player's nationality
Rating	Overall skill rating (45-99)
Market_Value	Estimated transfer value (formatted in Euro)
Scout_Report	AI-generated text describing playstyle and potential

---

🔍 Exploratory Data Analysis (EDA)

Below is the breakdown of the analysis performed on the synthetic data to verify its realism and uncover trends.

Step 1: Excel to CSV Conversion & Basic Check

This step processes the raw Excel generation output, converts it into a standardized CSV format for easier processing, and performs a sanity check on the row/column counts to ensure data integrity.

Step 2: IQR Bell Curve & Outlier Detection

We use the Interquartile Range (IQR) method to visualize the distribution of player ratings. This analysis identifies "Normal" players versus statistical "Outliers" (Superstars or low-tier players), visualizing them on a bell curve.

Step 3: Country Efficiency (Value vs. Quantity)

A scatter plot analyzing which nations produce the most value. It compares the quantity of players produced (X-axis) against their total combined market value (Y-axis), highlighting the most efficient footballing nations.

Step 4: League Power Rankings

A box plot visualization showing the skill spread across different leagues. The data is sorted by median player rating to determine the objective "Power Rankings" of each league, showing the range of talent from average starters to star players.

Step 5: Correlation Heatmap

A heatmap to identify what primarily drives a player's market value. We analyze the mathematical correlation between Age, Rating, and Market Value to visualize positive and negative relationships between variables.

Step 6: Feature Importance (Random Forest)

Using a Random Forest Regressor, we mathematically determine which features are the primary drivers of Market Value. We test two models: one including "Rating" (where skill dominates) and one without "Rating" (where Age typically becomes the dominant factor).

---

🏆 Phase 3: Embedding Model Tournament

To determine the best model for generating vector embeddings of the scouting reports, we conducted a "Tournament" benchmarking three distinct models. The goal was to find the optimal balance between inference speed and semantic quality.

The Contenders

1. Model A: paraphrase-MiniLM-L3-v2 (Fast / L3)
2. Model B: all-MiniLM-L6-v2 (Balanced / L6)
3. Model C: all-mpnet-base-v2 (Heavy / MPNet)

The Winning Logic (Score Calculation)

We did not simply choose the most accurate model (which is often too slow) or the fastest model (which is often inaccurate). Instead, we calculated a Weighted Efficiency Score using the following formula:

$$\text{Total Score}=(0.60\times \text{Quality Score})+(0.40\times \text{Speed Score})\backslash text\{Total\; Score\}\; =\; (0.60\; \backslash times\; \backslash text\{Quality\; Score\})\; +\; (0.40\; \backslash times\; \backslash text\{Speed\; Score\})$$

Where both Quality and Speed were normalized to a 0-100 scale before weighting. This ensured that Semantic Quality (60%) was prioritized slightly over pure Processing Speed (40%).

🥇 The Winner: Model A (Fast / L3)

Based on this formula, Model A achieved the highest combined score. It offered the best trade-off, allowing this app to run in real-time without lagging. The final embeddings for all 20,000 players were generated using this model.

⚙️ The App Architecture (Pipeline)

When you click "Scout" in the app, the system triggers the following 4-step pipeline to find and announce the perfect signing:

Step 1: Input Translation

The app accepts common football jargon and maps it to the dataset's standardized labels.

• User Input: "CF", "Left Back", "CDM"
• System Translation: "Striker", "Full Back", "CDM"

Step 2: Intelligent Filtering

We apply a custom clean_money_universal function to parse the user's budget and filter the 20,000-player dataset to only show realistic targets within financial constraints.

Step 3: Semantic Search (The Core)

Using the MiniLM-L3 embeddings chosen in our Tournament, the app vectorizes your text query and calculates the Cosine Similarity against the player database. This allows it to find players who match the style you described, not just the stats.

Step 4: Automated Content Generation

Once the perfect match is found, the system uses a Dynamic Template Engine to generate a realistic "Breaking News" tweet. This mimics the style of famous journalist Fabrizio Romano ("Here we go!"), creating instant social media assets for the transfer without requiring external API keys.

⚽ Scouter-AI Scouter-AI is an intelligent football scouting platform that allows users to find their ideal players using natural language. Instead of clicking through endless filters, you describe the player you want (e.g., "Young Brazilian striker under €50m"), and the AI handles the rest—complete with a "Here We Go!" transfer update from an AI-powered Fabrizio Romano.

🚀 Key Features Natural Language Scouting: A custom interpreter parses user queries for positions, ages, budgets, and nationalities.

Semantic Player Search: Uses vector embeddings to match the "vibe" or description of a player (like "creative playmaker" or "strong defender") to a real database.

Dynamic Player Cards: Generates FIFA-style visual cards for the top 3 matches, including scouting reports, ratings, and market values.

Fabrizio Romano Tweet Generator: Every search generates a unique, context-aware transfer update in the iconic style of the world's most famous football journalist.

🧠 AI Models & Technology Stack The intelligence of the app is powered by two primary AI implementations:

1. Large Language Model (LLM) for Tweet Generation To bring the "Fabrizio Romano" persona to life, the app utilizes the Meta-Llama-3-8B-Instruct model via the Hugging Face Inference API.

Role: Acts as a creative agent to transform raw player data (name, team, value) into a realistic, urgent transfer tweet.

Configuration: The model is prompted with a specific "system persona" to ensure it uses Romano's signature catchphrases ("Here we go!"), emojis (🚨, 🔴, 🔵), and insider tone.

2. Sentence Transformers for Semantic Search For the core "matching" logic, the app uses the paraphrase-MiniLM-L3-v2 model from the sentence-transformers library.

Role: It converts the user’s text query into a mathematical vector (embedding).

Process: The app compares the user's vector against pre-computed player embeddings using Cosine Similarity. This allows the app to understand that a "speedy attacker" is similar to a "pacey winger," even if those exact words aren't in the database.

3. Smart Logic Interpreter Regex-based Parsing: Custom-built logic to extract constraints like "U21" (age) or "50m" (budget) from messy human sentences.

Gibberish Detection: A safeguard mechanism that filters out nonsensical inputs, ensuring the AI only processes relevant football queries.

🛠️ How to Use Input: Type a description in the search bar (e.g., "Fast winger from Europe under 30m").

Process: The app filters the database by criteria, then ranks the remaining players using semantic similarity.

Output: View the top 3 prospects as interactive cards and read the "exclusive" transfer news generated by the Llama-3 model.

🔗 Links & Resources

• Dataset: Hugging Face Dataset Card
• Project Notebook: See the Showcase_Notebook.ipynb file in this repository for the full code analysis.