newlantern-deberta-sk

DeBERTa-v3-large fine-tuned for radiology prior study relevance classification. Given a current and a prior radiology study description, predicts whether the prior is relevant to the current read.

Built for the New Lantern challenge: relevant-priors-v1.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("sh3hryarkhan/newlantern-deberta-sk")
model = AutoModelForSequenceClassification.from_pretrained("sh3hryarkhan/newlantern-deberta-sk")
model.eval()

text_a = "delta: 365d (<= 1y) | cur: CT HEAD WITHOUT CONTRAST | norm: CT/HEAD/BILATERAL/WITHOUT"
text_b = "prior: MRI BRAIN ROUTINE | norm: MRI/HEAD/BILATERAL/WITHOUT"

enc = tokenizer(text_a, text_b, truncation=True, max_length=128, return_tensors="pt")
with torch.no_grad():
    prob = torch.softmax(model(**enc).logits, dim=-1)[0, 1].item()

print(f"P(relevant) = {prob:.3f}")

Label 1 = relevant, 0 = not relevant. Decision threshold: 0.5.

Input Format

text_a: delta: {days}d ({bucket}) | cur: {current_description} | norm: {mod}/{region}/{lat}/{con}
text_b: prior: {prior_description} | norm: {mod}/{region}/{lat}/{con}

The norm field is a 4-tuple (modality, region, laterality, contrast) parsed from the raw description string. The delta bucket maps days to one of: same day, <= 1m, <= 3m, <= 6m, <= 1y, <= 2y, <= 5y, > 5y.

Training

• Base model: microsoft/deberta-v3-large
• Task: binary sequence classification (relevant / not relevant)
• Data: public split of the New Lantern relevant-priors-v1 challenge (~13k labeled pairs)
• Epochs: 4 (early stopped; epoch 5 overfit)
• Phase 1 (epochs 1-2): standard cross-entropy training; AdamW lr=2e-5, cosine schedule, fp16
• Phase 2 (epochs 3-4): hard-negative mining via WeightedRandomSampler (3x weight on misclassified samples)
• Hardware: A100
• Best val accuracy: 96.44%

System Context

This model is Tier 2 in a three-tier cascade. Tier 1 is a lookup table over seen (cur, prior) description pairs with Laplace smoothing; the encoder only fires when the lookup abstains (novel pairs). In production on the full public split, ~5-10% of pairs reach the encoder. See the full system at sh3hryarkhan/newlantern-deberta-sk.