KevinDavidHayes/regressor-local-ratio

Long-Context Attention Regressor (Local Ratio)

Predicts how much a text sample uses long-range attention vs local attention when processed by a transformer LLM.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained("KevinDavidHayes/regressor-local-ratio")
tokenizer = AutoTokenizer.from_pretrained("KevinDavidHayes/regressor-local-ratio")

text = "Your text here..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=8192)

with torch.no_grad():
    score = model(**inputs).logits.item()

# Higher score = more long-range attention patterns

Training

• Base model: ModernBERT-base (8K context)
• Target: Inverted local_ratio (fraction of attention outside 128-token window)
• Labels: Generated using Qwen2.5-7B-Instruct attention analysis at layer 14

Metrics

Metric	Value
Test R²	See regressor_metadata.json
Test MAE	See regressor_metadata.json

Citation

Part of research on attention-based data filtering for long-context pretraining.