Tanushreenerella/closira-multitask

Multi-task DistilBERT (distilbert-base-uncased) for the Closira IT helpdesk router: one encoder, an intent head and an emotion head, exported to ONNX with dynamic int8 quantization. Served with onnxruntime + tokenizers (no torch). baseline/ holds the TF-IDF + LogReg fallback.

• intents: information_request, issue_report, action_request, complaint, feedback, appreciation, chit_chat
• emotions: angry, frustrated, confused, sad, happy, neutral
• inputs: input_ids, attention_mask (int64, max length 128); outputs: intent_logits, emotion_logits

Eval (PyTorch model)	Accuracy	Macro-F1
intent test	0.9767	0.9819
emotion test	0.8217	0.7515

int8 vs PyTorch prediction agreement: intent 0.9898, emotion 0.9217.

Training data: Banking77, CLINC150, SNIPS, TWCS (keyword weak labels), GoEmotions. See backend/ml/results/README.md in the Closira repo for evaluation details and caveats.