The goal of this project is to fine-tune microsoft/SportsBERT to extract PLAYER, INJURY, STATUS, and TEAM from sports news using a Weak Supervision approach.

Intended Use

The model is designed to extract the following entities from sports news:

PLAYER: Name of the injured player.
INJURY: Type of injury (e.g., "hamstring", "concussion").
STATUS: Injury status (e.g., "questionable", "out", "placed on IR").
TEAM: Team name (e.g., "Packers", "New York Giants").

Usage

python

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

model_id = "maxo99/sports-injury-ner"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

nlp = pipeline("ner", model=model, tokenizer=tokenizer, aggregation_strategy="simple")

text = "Packers QB Aaron Rodgers is questionable with a toe injury."
results = nlp(text)
print(results)

Performance

Metric	Score
F1 Score	94.98%
Precision	95.41%
Recall	94.55%
Accuracy	97.06%

Evaluated: Dec 2025.

Training Data

The model was fine-tuned using a Weak Supervision + Active Learning approach:

Silver Label Generation: Initial labels generated using bert-large-NER (for Player/Team) and keyword matching (for Injury/Status).
Active Learning:
- Uncertainty Sampling: "Least Confidence" scoring to identify confusing examples.
- Conflict Detection: Flagging examples where keywords disagree with model predictions.
- Human-in-the-loop: Manual validation of high-value candidates to create a "Gold Standard".
Fine-tuning: microsoft/SportsBERT fine-tuned on the combined Silver + Gold dataset.

Limitations

Entity Resolution Priority: The data generation script resolves overlapping entities using a strict priority: Metadata > Keywords > BERT NER.
Domain Specificity: The model is trained on a specific dataset of NFL/ESPN reports and may not generalize well to other sports or writing styles.

Development

For instructions on how to reproduce the training pipeline, run tests, or contribute, please see docs/development.md.

Quick Start (DVC)

This project uses DVC to manage the data pipeline.

bash

# 1. Install dependencies
uv sync

# 2. Run the full pipeline (Data Gen -> Validation -> Training)
dvc repro train_model