---
license: cc-by-4.0
library_name: scikit-learn
tags:
  - hackathon
  - tabular-classification
  - archetype
  - alphahack
pipeline_tag: tabular-classification
---

# AlphaHack Model 1 — Event Regime Classifier

A `GradientBoostingClassifier` (with a logistic-regression fallback,
LabelEncoder, and StandardScaler bundled in the same pickle) that
predicts which **winner archetype** dominates a given hackathon event.

Companion: [Model 2 — winner predictor](https://huggingface.co/xenosaac/alphahack-models/tree/main/model2-winner-predictor)

## Model description

Given event-level features (host type, judge composition, prize pool,
duration, theme keywords, criteria text, etc.), the classifier predicts
which of 5 archetypes describes the event's prior winners:

| Archetype | Train rows |
|---|---|
| `tech_showoff` | 1,051 |
| `empathy_play` | 825 |
| `scrappy_utility` | 761 |
| `hype_surfer` | 717 |
| `narrative_master` | 658 |

Use the model to inform **idea-generation strategy** before a hackathon
— not to rank individual project submissions (use Model 2 for that).

## Training data

The model was fit on **4,012 event-level archetype labels** derived from
**23,785 winning-project rows** across **101,682 total projects** in
the [`xenosaac/alphahack-devpost`](https://huggingface.co/datasets/xenosaac/alphahack-devpost)
dataset. Source feature parquet:
`data/merged/alphahack_features_v7.parquet` (151 columns post-PII-scrub).

## Metrics

5-fold cross-validation, GroupKFold by `event_id` (no event appears in
both train and test of the same fold).

| Metric | GBC | LR baseline | Majority baseline |
|---|---|---|---|
| Top-1 accuracy | **0.381 ± 0.012** | 0.275 | 0.262 |
| Top-3 accuracy | **0.804 ± 0.010** | — | — |
| Train accuracy | 0.783 | — | — |

GBC delivers a **1.45× lift** over the majority baseline on top-1
accuracy and a **2.92×** lift over majority for the practically more
useful top-3 accuracy (which is what feeds the strategy engine's
portfolio prompting).

## Top features (GBC importance)

| Feature | Importance |
|---|---|
| `A06_total_submissions` | 0.291 |
| `A11_theme_keywords` | 0.075 |
| `A09_num_prize_categories` | 0.068 |
| `A05_prize_pool_usd` | 0.064 |

## Loading the model

```python
import joblib

bundle = joblib.load("regime_classifier.pkl")
gbc = bundle["gbc"]                # primary classifier
lr = bundle["lr"]                  # logistic-regression baseline
le = bundle["le"]                  # LabelEncoder for archetype names
scaler = bundle["scaler"]          # StandardScaler (event features)
feature_cols = bundle["feature_cols"]  # list of 32 input column names

# Score a new event
import numpy as np
event_features_dict = {...}   # build from your crawled event
X_raw = np.array([[event_features_dict[c] for c in feature_cols]])
X = scaler.transform(X_raw)
proba = gbc.predict_proba(X)[0]
top3_idx = proba.argsort()[::-1][:3]
top3_archetypes = le.inverse_transform(top3_idx)
print(list(zip(top3_archetypes, proba[top3_idx])))
```

## Reproducing this artifact

The full training pipeline is in the open-source companion repo:

```bash
pip install hackalpha
hackalpha train-model1 \
  --features data/merged/alphahack_features_v7.parquet \
  --model-out data/models/regime_classifier.pkl \
  --metrics-out data/research/model1_training_metrics.json \
  --report-out  data/research/model1_archetype_report.json
```

The training metrics (`model1_training_metrics.json`) and label
distribution (`model1_archetype_report.json`) are included in this
HF directory.

## Known failure modes

- Top-1 accuracy of 38% is well below human-expert level. The
  product-relevant metric is top-3 (80%), used to prompt a multi-idea
  portfolio rather than commit to one bet.
- 3 test years available (2024–2026) is not enough for tight CIs on
  Model 1's archetype labels.
- The label assignment is **heuristic-based** (a project is "tech_showoff"
  if its rubric scores match a tech-showoff signature), not adjudicated
  by humans. Label noise is real.

## Limitations

- Trained only on Devpost-hosted, English-language hackathons.
- In-person and non-Devpost events: performance unknown.
- Companion model 2 had a **prospective trial in April 2026 that did
  not produce a prize**. Use both models as research artifacts, not
  as a guaranteed winning recipe.

## License

CC BY 4.0.