JamieYCR/text-emotion-detector-final

Text Emotion Detector

This is a multi-label emotion classifier for conversational English text. It is based on microsoft/deberta-v3-large, adapted on Empathetic Dialogues, fine-tuned on GoEmotions, and then optionally tuned on hard examples for subtle distress/self-dismissal style inputs.

The model predicts 28 GoEmotions labels and is intended to return the top emotions for a piece of text rather than a single exclusive class.

Install

pip install torch transformers huggingface_hub

Basic Usage

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo = "JamieYCR/text-emotion-detector-final"

tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)
model.eval()

text = "I can't believe she actually said yes!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits[0]
    probs = torch.sigmoid(logits)

top_k = 3
top = probs.argsort(descending=True)[:top_k]

for idx in top:
    label = model.config.id2label[int(idx)]
    score = float(probs[idx])
    print(f"{label}: {score:.4f}")

Calibrated Probabilities

This repo includes temperature.json, which was fitted during calibration. Plain transformers loading does not apply it automatically. To use calibrated probabilities, divide logits by the temperature before applying sigmoid:

import json
import torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo = "JamieYCR/text-emotion-detector-final"

tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo)
model.eval()

temperature_path = hf_hub_download(repo_id=repo, filename="temperature.json")
with open(temperature_path, "r", encoding="utf-8") as f:
    temperature = float(json.load(f)["temperature"])

text = "I'm so tired of pretending everything is fine."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits[0]
    probs = torch.sigmoid(logits / temperature)

top = probs.argsort(descending=True)[:3]
results = [
    (model.config.id2label[int(i)], float(probs[i]))
    for i in top
]

print(results)

Batch Inference

texts = [
    "Good morning, I hope you have a wonderful day.",
    "Ugh, fine. Whatever.",
]

inputs = tokenizer(
    texts,
    return_tensors="pt",
    padding=True,
    truncation=True,
    max_length=128,
)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.sigmoid(logits)

for text, row in zip(texts, probs):
    top = row.argsort(descending=True)[:3]
    print(text)
    for idx in top:
        print(f"  {model.config.id2label[int(idx)]}: {float(row[idx]):.4f}")

Labels

The model uses the GoEmotions label set:

admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise, neutral.

Notes

• This is a multi-label classifier, so use torch.sigmoid, not softmax.
• The model was trained for top-k emotion ranking. For most applications, inspect the top 3-5 labels.
• The uploaded temperature.json can be used for calibrated probabilities.
• Empathetic Dialogues is licensed for non-commercial use, so treat this model as non-commercial unless retrained without that data.

Example Repository ID

repo = "JamieYCR/text-emotion-detector-final"