منصوبے کے بارے میں

ایم ایم اے فائٹ پریڈکشن ایک پایتھون ریسرچ اور سروسنگ پروجیکٹ ہے جو یو ایف سی فائٹ وینرز کی پیش گوئی کرتا ہے، وہت آف victority اور فنیس راؤنڈ۔ یہ ایلو ریٹنگ سسٹم، ایک گریڈینٹ بوُسٹڈ (ایکس جی بوُسٹ) انسامبل، ایک ملٹی ٹاسک نیورل نیٹ ورک انسامبل، اور دو کا ایک equal-weight blend استعمال کرتا ہے جو deployed scorer ہے۔ ہر candidate کو expanding-window walk-forward evaluation کے ذریعے 2018-2026 میں judged کیا جاتا ہے، اور winning scorer کو calibrated کیا جاتا ہے held-out data کے بعد blend average کے۔ Reproducible pipeline Quickstart virtual environment بنتا ہے، package install کرتا ہے dev extras کے ساتھ، پھر چار scripts چلاتا ہے order میں: download_data.py (Kaggle UFC dataset into data/raw/)، make_dataset.py (clean parquet into data/processed/)، build_ratings.py (tune and build Elo ratings)، followed by pytest۔ ایک Streamlit app (streamlit run app.py) user کو pick کرنے دیتا ہے دو fighters اور returns win probability with uncertainty، method اور finish-round tendencies، اور top contributing factors کا breakdown۔ ایک documented aim ہے byte-for-byte reproducibility of the whole pipeline۔ Feature engineering and leak control Features are assembled from named blocks (base، external، trajectory، notice، context، opponent_adjusted) in src/mma/feature_blocks.py، producing a table of roughly 11 thousand fights by 87 columns۔ Features are built inside chronological accumulators so each value is strictly pre-fight، اور ایک truncation-invariance test استعمال ہوتا ہے check کرنے کے کہ no feature can see the future۔ Blocks were measured against a pre-registered accuracy bar: one (external، covering pre-UFC career data) cleared it on its own، while others ship only as part of the table the blend was scored on، or do not ship at all۔ Project documents a selection leak found in per-corner missingness flags، which was removed from both model matrices while staying in the table for evaluation slices۔ Several columns are explicitly excluded from the model inputs as leak guards، اور ایک provenance sidecar file records per row کہ آیا came from the primary or secondary source۔ Data sources Records are assembled as a union of a maintained Kaggle mirror (the primary، reconciled source) اور Greco1899/scrape_ufc_stats CSV publication (used strictly as data، fetched over HTTPS، with no GPL-licensed code vendored)۔ On overlapping fights the primary row wins؛ merge is gated on winner agreement between sources اور is fail-soft، degrading to a primary-only build with a warning۔ Repository reports provenance، integrity checks اور regression guards around this merge۔ Self-updating and prospective evaluation Weekly GitHub Action refreshes dataset اور rebuilds artifacts، committing them automatically۔ Upcoming UFC events are predicted اور committed to git before they happen اور graded afterwards، giving a prospective track record rather than retrospective-only reporting۔ Harness reports for rejected feature blocks اور pre-registered experiment decision files are committed under models/walkforward/، so negative results are retained۔ Project is written in Python (3.10+) اور released under MIT license۔ Design اور phase-plan documents live under docs/۔