patelkrish2727/jarvis-email

Jarvis Email Intelligence

Custom multi-head email classifier on microsoft/deberta-v3-small.

Heads:

• category (11): finance, government, newsletters, other, personal, promotions, security, socials, spam, verification_codes, work
• intent (10): financial_transaction, government_notice, information_update, meeting_request, other, promotional_offer, security_alert, social_notification, task_action, verification_code
• actionability (5): needs_click, needs_payment, needs_reply, needs_scheduling, none_fyi
• threat_verdict (5): fraud_bec, impersonated, legitimate, phishing, suspicious
• attack_vector (6): credential_harvesting, fake_invoice, malicious_link, none, payment_diversion, social_engineering
• binary: requires_response, display_name_mismatch, urgency_manipulation
• scores 0-1: importance, urgency, threat, confidence

Files

• pytorch_model.bin — full state_dict (540MB, fp32)
• modeling_jarvis.py — model class JarvisEmailIntelligence
• config.json — dims + label lists
• label_encoders.pkl — sklearn LabelEncoders (sklearn 1.6.1)
• label_mappings.json — same labels as JSON

Quick use

import pickle, torch
from transformers import AutoTokenizer
from modeling_jarvis import JarvisEmailIntelligence

REPO = "patelkrish2727/jarvis-email"
tok = AutoTokenizer.from_pretrained("microsoft/deberta-v3-small")
encoders = pickle.load(open("label_encoders.pkl","rb"))

model = JarvisEmailIntelligence(
  model_name="microsoft/deberta-v3-small",
  n_category=len(encoders['category'].classes_),
  n_intent=len(encoders['intent'].classes_),
  n_actionability=len(encoders['actionability'].classes_),
  n_threat_verdict=len(encoders['threat_verdict'].classes_),
  n_attack_vector=len(encoders['attack_vector'].classes_),
)
state = torch.load("pytorch_model.bin", map_location="cpu")
model.load_state_dict(state); model.eval()

text = "Subject: Hello\n\nBody: test"
e = tok(text, truncation=True, padding='max_length', max_length=384, return_tensors='pt')
with torch.no_grad():
    print(model(e['input_ids'], e['attention_mask']))

Max length: 384. Input format: Subject: {subject}\n\nBody: {body}.