README.md

---
license: other
pretty_name: MMA AI database and model artifacts
size_categories:
- 1M<n<10M
task_categories:
- tabular-classification
tags:
- mma
- ufc
- postgresql
- autogluon
- sports-analytics
---

# MMA AI Dataset Artifacts

This dataset contains the database dumps and runtime artifacts needed to
reproduce the [`mma-ai`](https://github.com/DanMcInerney/mma-ai) workflow.

The current release was refreshed on 2026-08-16 from PostgreSQL 18.1 and uses
the accepted recency-weighted v8 hybrid model. Exact sizes, hashes, database
metadata, model lineage, and evaluation boundaries are in `manifest.json`.

## Contents

- `dumps/mma-ai.postgres-custom` — custom-format PostgreSQL dump containing the
  main `features` schema used by `DATABASE_URL`.
- `dumps/odds.postgres-custom` — custom-format PostgreSQL dump containing
  `bestfightodds.bfo`, used by `ODDS_DATABASE_URL`.
- `processed/training_data.csv` — generated win-model training data.
- `processed/training_data_dec.csv` — generated decision-model training data.
- `processed/prediction_data.csv` — generated prediction feature data.
- `models/ag-20260815_090928-win-hybrid.tar.gz` — accepted AutoGluon weighted-v8
  hybrid win model.
- `manifest.json` — authoritative sizes, SHA-256 hashes, database metadata, and
  source/model lineage.

The dumps use PostgreSQL custom archive format with gzip compression.

## Model evaluation boundary

The published model uses 40 v8 features, recency weighting, a chronological
split, and the hybrid candidate set. Its saved non-FULL ensemble is
`1.0 * Mitra`; the archive contains 22 AutoGluon model nodes including FULL
variants.

On its exposed 460-fight chronological tuning partition it scored 309/460
(67.17% accuracy, 0.61319 log loss). This is a selection-biased tuning result,
not an untouched holdout. The separately replayed, event-grouped chronological
development estimate across 2022–2025 was 726/1,108 (65.52% accuracy, 0.61960
log loss). See the companion repository's
`research/2026-08-16-top10-mma-experiments/` evidence for the full provenance.

## Restore databases

Create local databases:

```bash
createdb -U postgres mma-ai
createdb -U postgres odds
```

Restore the dumps using the host/port for your own PostgreSQL service:

```bash
pg_restore --clean --if-exists --no-owner --jobs 4 \
  --dbname "postgresql://postgres@localhost:5432/mma-ai" \
  dumps/mma-ai.postgres-custom

pg_restore --clean --if-exists --no-owner --jobs 4 \
  --dbname "postgresql://postgres@localhost:5432/odds" \
  dumps/odds.postgres-custom
```

## Use the pretrained model

Extract the model into the code repository:

```bash
mkdir -p AutogluonModels
tar -xzf models/ag-20260815_090928-win-hybrid.tar.gz -C AutogluonModels
mkdir -p data
cp processed/training_data.csv data/training_data.csv
cp processed/prediction_data.csv data/prediction_data.csv
```

Then run:

```bash
uv run python predict.py \
  --model-path AutogluonModels/ag-20260815_090928-win-hybrid \
  --prediction-data-csv data/prediction_data.csv \
  --training-data-csv data/training_data.csv \
  --no-shap
```

## Rebuild and retrain instead

After restoring both databases, rebuild the generated data and train with the
companion repository's versioned profiles. The processed CSVs are included so
users can skip that rebuild for the common prediction workflow.