ethnmcl/mvp-router-roberta

MVP Router – RoBERTa (Text-Only)

A high-precision text routing model built on RoBERTa for Q/A applications.

This model classifies a user’s question into a single routing label based only on the text, and supports abstention (NEEDS_CLARIFICATION) when confidence is below a learned threshold.

---

🎯 Purpose

This model is not a chatbot.

It is designed to act as a router that determines where a question should go inside an application, especially when different user roles exist.

Incorrect routing — particularly for Admin or Parent queries — is treated as a failure case.

---

🏷️ Labels

The model predicts exactly one of the following:

Label	Description
CIC	Internal organization member (check-ins, meetings, processes)
Admin	Organization controller (permissions, audits, data correction)
Request	End-user feature requests (recaps, MVT points, stories, coaching)
Parent	Guardian asking about a child’s activity
Other	Out-of-scope or unrelated questions

---

🧠 Key Features

• Text-only inference (no metadata, headers, or auth context)
• Calibrated probabilities (temperature scaling)
• Per-label confidence thresholds
• Abstention support
  → returns NEEDS_CLARIFICATION when confidence is too low
• Safety-first routing (Admin & Parent precision prioritized)

---

📦 Files in This Repo

File	Purpose
pytorch_model.bin / model.safetensors	Trained RoBERTa weights
config.json	Model configuration
tokenizer.json / vocab files	Tokenizer
thresholds.json	Learned confidence thresholds per label
label_map.json (if present)	Explicit label ↔ id mapping
README.md	This file

---

🚀 Quick Inference Example

import json
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification

MODEL_ID = "ethnmcl/mvp-router-roberta"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID).to(DEVICE)
model.eval()

with open("thresholds.json", "r") as f:
    thresholds = json.load(f)

id2label = model.config.id2label

def route(text: str):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(DEVICE)

    with torch.no_grad():
        logits = model(**inputs).logits
        probs = F.softmax(logits, dim=-1).squeeze()

    pred_id = int(torch.argmax(probs))
    pred_label = id2label[pred_id]
    confidence = float(probs[pred_id])

    if confidence < thresholds[pred_label]:
        return "NEEDS_CLARIFICATION", confidence

    return pred_label, confidence

print(route("Show me my MVT points from yesterday."))