Fintacorp55/football-llm-q4

⚽ Football Coaching LLM — Qwen2 7B (DPO v2)

A holistic football coaching assistant fine-tuned on Qwen2-7B using supervised fine-tuning (SFT) + Constitutional AI preference learning (DPO). Built on a RAG knowledge base of ~500M tokens spanning 11 domains, 66 subthemes, and 460+ football concepts.

Available Models

This repository contains multiple model formats to suit different deployment scenarios:

Model Details

Capabilities

• —Session planning & drill design — Generate structured training sessions across all age groups (U8 to Senior), with progressions, coaching points, and time breakdowns.
• —Tactical analysis — Formations, pressing structures, build-up play, transition patterns, and set-piece strategies grounded in the RAG knowledge base.
• —Periodization — Season planning, mesocycle/microcycle structuring, and load management frameworks.
• —Player development — Age-appropriate coaching pathways, individual development plans, and talent identification principles.
• —Coaching cues & feedback — Specific verbal and visual cues for technique correction and tactical communication.
• —Holistic coaching methodology — Integrates tactical, physical, mental, and social dimensions of player development.

Usage

Option 1: LoRA Adapter (Transformers + PEFT)

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model = "unsloth/qwen2-7b-bnb-4bit"
adapter_path = "Fintacorp55/football-llm-q4"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(adapter_path)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    quantization_config=bnb_config,
    device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_path)

messages = [{"role": "user", "content": "Design a pressing drill for U16 players."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Option 2: GGUF via llama.cpp (Python)

from llama_cpp import Llama

llm = Llama(
    model_path="./football-llm-q4_K_M.gguf",
    n_ctx=4096,
    n_threads=8,
    n_gpu_layers=35  # Adjust for your hardware; set to 0 for CPU-only
)

output = llm(
    "Explain the principles of positional play in a 4-3-3 formation.",
    max_tokens=512,
    temperature=0.7,
)
print(output["choices"][0]["text"])

Option 3: Ollama

# Import the GGUF directly
ollama create football-coach -f Modelfile
ollama run football-coach

Option 4: LM Studio

Download football-llm-q4_K_M.gguf (4.68 GB) or football-llm-q4-f16.gguf (15.2 GB) and load directly in LM Studio.

Training Pipeline

RAG Knowledge Base (~500M tokens)
        │
        ▼
┌─────────────────────┐
│  1. Data Cleaning    │  10,000 filtered examples from RAG corpus
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  2. Synthetic QA     │  3,002 Claude-generated football Q&A pairs
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  3. SFT v2           │  Fine-tuned on 13,002 merged examples
│                       │  3 epochs · LoRA rank 32 · alpha 64
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  4. CAI Pairs         │  560 Constitutional AI preference pairs
│                       │  Generated via Claude API
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  5. DPO v2            │  Preference learning · beta=0.1
│                       │  Reward accuracy: 100%
└────────┬────────────┘
         │
         ▼
┌─────────────────────┐
│  6. GGUF Export       │  Q4_K_M (4.68 GB) + F16 (15.2 GB)
│                       │  Merged LoRA → full model → quantized
└─────────────────────┘

Data Sources

The RAG knowledge base that produced the training data draws from three pillars:

• —FBref Analytics — Player and team advanced metrics (xG, xA, progressive actions, defensive stats) across the European Big 5 leagues.
• —Cocoon Tactical Articles — In-depth tactical analysis: formations, pressing systems, transition play, positional structures, coaching methodologies.
• —Curated Football Corpus — Coaching manuals, periodization frameworks, player development literature, and match analysis content.

Knowledge Map

The underlying RAG structure is visualized as an interactive mindmap at llm.fintalab.com/knowledge_map:

• —11 Domains — High-level categories spanning the full football intelligence landscape.
• —66 Subthemes — Granular topic areas within each domain.
• —460+ Concepts — Specific knowledge nodes extracted from the RAG corpus.
• —Council Agent System — Select domain experts from the map to assemble a multi-agent "council." The system generates tailored prompts for focused, multi-perspective analysis.

You can also explore the RAG conversationally via the Football RAG on NotebookLM.

Part of the Finta Ecosystem

This model is one component of a broader sports intelligence platform at fintalab.com — a 12-instrument ecosystem that decomposes athletic performance into measurable dimensions (tactical, predictive, biomechanical, cognitive) across football, tennis, and general sports coaching. The LLM and Knowledge Map are instruments #11 and #12 in that toolkit.

Limitations

• —Domain scope — Optimized for football/soccer coaching contexts. Performance on non-football queries will default to the base Qwen2 model's capabilities.
• —Expert validation — Responses may require professional review for elite-level tactical or physical preparation decisions.
• —League coverage — FBref data skews toward the European Big 5 (Premier League, La Liga, Bundesliga, Serie A, Ligue 1). Lower divisions and non-European leagues have less coverage.
• —Knowledge cutoff — Based on training data collection date; no live or real-time match data.
• —Quantization trade-offs — Q4KM reduces size and improves speed at the cost of some precision vs. F16.
• —Language — Primarily English. Some French-language tactical content from Ligue 1/Ligue 2 sources.

Ethical Considerations

• —Intended for educational, analytical, and coaching purposes.
• —Should not be used for gambling, match manipulation, or sports integrity violations.
• —Player data sourced from publicly available statistics (FBref) and published analyses.
• —Users should exercise critical judgment when applying AI-generated coaching recommendations.

Citation

@misc{finta_football_llm_2025,
  title={Football Coaching LLM: A Domain-Specific Fine-Tuned Model with RAG Knowledge Base for Football Intelligence},
  author={Johnson, Dylan},
  year={2025},
  url={https://huggingface.co/Fintacorp55/football-llm-q4},
  note={Finta Sports / FintaLab}
}

Contact

• —Website — fintalab.com
• —LinkedIn — Dylan Johnson
• —Knowledge Map — llm.fintalab.com/knowledge_map
• —RAG Knowledge Base — NotebookLM

Questions, feedback, or collaboration ideas? Let's talk on LinkedIn.