sentinelai-bert-filter

LoRA Configuration:
  r: 8
  lora_alpha: 16
  lora_dropout: 0.1
  target_modules: ["query", "value"]
  task_type: FEATURE_EXTRACTION

Training:
  epochs: 3
  batch_size: 16
  learning_rate: 3e-4
  optimizer: AdamW
  scheduler: Linear warmup + decay
  max_sequence_length: 128
  loss_function: CrossEntropyLoss (category + severity summed)

import torch
from services.model_factory import load_production_model

# Inference
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# This will:

# 1. Initialise DualHeadBERTClassifier
# 2. Apply LoRA adapters
# 3. Check for 'dual_head_classifier.pt' locally
# 4. If missing, download latest from OguzhanKOG/sentinelai-bert-filter
# 5. Load trained weights and return model in eval mode

model = load_production_model(device=device)

# Model is ready for inference
message = "I'm completely overwhelmed with work and can't sleep anymore"
# ... standard tokenization using config.MODEL_NAME ...