---
language:
- en
license: mit
library_name: founder-game-classifier
tags:
- text-classification
- founders
- content-analysis
- sentence-transformers
datasets:
- custom
metrics:
- accuracy
pipeline_tag: text-classification
---

# Founder Game Classifier

A trained classifier that identifies which of **6 founder games** a piece of content belongs to.

## Model Description

This model classifies text content into one of six "founder games" - patterns of communication and content creation common among founders, creators, and thought leaders.

### The 6 Games

| Game | Name | Description |
|------|------|-------------|
| **G1** | Identity/Canon | Recruiting into identity, lineage, belonging, status, canon formation |
| **G2** | Ideas/Play Mining | Extracting reusable plays, tactics, heuristics; "do this / steal this" |
| **G3** | Models/Understanding | Building mental models, frameworks, mechanisms, explanations |
| **G4** | Performance/Competition | Winning, dominance, execution, metrics, endurance, zero-sum edges |
| **G5** | Meaning/Therapy | Healing, values, emotional processing, personal transformation |
| **G6** | Network/Coordination | Community building, protocols, collaboration, collective action |

## Usage

### Installation

```bash
pip install founder-game-classifier
```

### Basic Usage

```python
from founder_game_classifier import GameClassifier

# Load the model (downloads from Hub on first use)
classifier = GameClassifier.from_pretrained("leoguinan/founder-game-classifier")

# Classify a single text
result = classifier.predict("Here's a tactic you can steal for your next launch...")

print(result["primary_game"])      # "G2"
print(result["confidence"])        # 0.72
print(result["probabilities"])     # {"G1": 0.05, "G2": 0.72, "G3": 0.10, ...}
```

### Batch Classification

```python
texts = [
    "Here's the mental model I use for thinking about systems...",
    "Join our community of builders who are changing the world...",
    "I tried 47 different tactics. Here's what actually worked...",
]

results = classifier.predict_batch(texts)

for text, result in zip(texts, results):
    print(f"{result['primary_game']}: {text[:50]}...")
```

### Get Aggregate Signature

Useful for analyzing a corpus of content:

```python
texts = load_my_blog_posts()  # List of strings
signature = classifier.get_game_signature(texts)

print(signature)
# {'G1': 0.05, 'G2': 0.42, 'G3': 0.18, 'G4': 0.20, 'G5': 0.08, 'G6': 0.07}
```

## Model Architecture

- **Embedding Model**: `all-MiniLM-L6-v2` (384 dimensions)
- **Classifier**: Logistic Regression (sklearn)
- **Manifold System**: Mahalanobis distance to game centroids (optional)

## Training Data

The model was trained on labeled founder content spanning:
- Podcast transcripts
- Blog posts
- Twitter threads
- Newsletter content

Training used a multi-stage pipeline:
1. Text chunking and span extraction
2. LLM-assisted labeling with human verification
3. Embedding generation
4. Classifier training with cross-validation

## Performance

Validated on held-out test set:

| Metric | Score |
|--------|-------|
| Accuracy | 0.78 |
| Macro F1 | 0.74 |
| Top-2 Accuracy | 0.91 |

The model performs best on clear examples of each game and may show lower confidence on boundary cases or mixed content.

## Limitations

- Trained primarily on English content from tech/startup domain
- May not generalize well to non-business contexts
- Short texts (<50 words) may have lower accuracy
- Cultural and domain biases from training data

## Citation

```bibtex
@misc{guinan2024foundergameclassifier,
  title={Founder Game Classifier},
  author={Leo Guinan},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/leoguinan/founder-game-classifier}
}
```

## License

MIT License - free for commercial and non-commercial use.

## Files

- `classifier.pkl` - Trained LogisticRegression model (19KB)
- `label_encoder.pkl` - Label encoder for game classes (375B)
- `metadata.json` - Model metadata and configuration (143B)
- `game_manifolds.json` - Manifold centroids and covariances for geometric analysis (29MB)