KevinDavidHayes/regressor-composite

Long-Context Attention Regressor (Composite)

Predicts a composite score combining multiple attention metrics to identify text that benefits from long context.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

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

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 = text benefits more from long-range attention

Training

• Base model: ModernBERT-base (8K context)
• Target: Weighted combination: 0.2 * mean_distance + 0.4 * inv_local_ratio + 0.4 * entropy
• Labels: Generated using Qwen2.5-7B-Instruct attention analysis at layer 14

Why Composite?

Cross-context correlation analysis showed:

• mean_distance: r=0.71 (4K→32K)
• local_ratio: r=0.92
• entropy: r=0.92

The composite weights metrics by their cross-context stability.

Citation

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