README.md

---
license: mit
base_model: Qwen/Qwen2.5-3B-Instruct
language:
- en
pipeline_tag: text-generation
tags:
- football
- sports
- json-extraction
- gguf
- lora
- qwen2.5
library_name: llama-cpp
---

# forecast-extractor

A fine-tuned version of [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) 
for extracting structured JSON from football prediction messages (e.g. Telegram tip channels).

## What it does

Given a raw football prediction message, it returns a structured JSON array:
```json
[
  {
    "league": "La Liga",
    "team_1": "Real Madrid",
    "team_2": "Barcelona",
    "prediction": "1X",
    "date": "25/03/2026",
    "odds": 1.42
  }
]
```

Handles:
- Single and multi-tip messages (up to 4 tips)
- Bold unicode text (Telegram formatting)
- Missing fields → null
- Varied formats, emojis, noise

## Models

| File | Size | Description |
|---|---|---|
| `football-extractor-q4.gguf` | 1.8GB | Q4_K_M quantized — recommended |
| `football-extractor-f16.gguf` | 5.8GB | Full f16 precision |

## Quick start

### With llama-cpp-python (recommended)
```python
from llama_cpp import Llama
import json

llm = Llama(model_path="football-extractor-q4.gguf", n_ctx=2048, n_gpu_layers=-1)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "Extract structured data and return ONLY a valid JSON array. Keys: league, team_1, team_2, prediction, date, odds. Use null for missing fields."},
        {"role": "user", "content": "YOUR TIP TEXT HERE"}
    ],
    temperature=0.0,
    max_tokens=512,
)
print(json.loads(response["choices"][0]["message"]["content"]))
```

### With Ollama
```bash
ollama pull philippotiger/forecast-extractor
ollama run philippotiger/forecast-extractor
```

## Training details

- **Base model:** Qwen/Qwen2.5-3B-Instruct
- **Method:** QLoRA (4-bit NF4) with LoRA r=8
- **Dataset:** 300 synthetic examples generated from real team data
  - 70% single-tip, 30% multi-tip (2-4 events)
  - 10 message templates with emoji injection, typos, missing fields
- **Epochs:** 3
- **Final val loss:** ~0.24

## Intended use

Parsing football prediction messages from Telegram channels or similar 
sources into structured data for further processing or storage.