You can deploy this model and start earning money today!

RenDeniz/
emotion-detective-v2

Name: emotion-detective-v2
Availability: InStock

safetensors

distilbert

Views

No views yet

Emotion Detective v2 — DistilBERT multi-label emotion classifier

DistilBERT (distilbert-base-uncased) fine-tuned on GoEmotions for multi-label classification across 28 emotion labels (27 emotions + neutral).

👉 Live demo: https://huggingface.co/spaces/RenDeniz/emotion-detective

Intended use

Portfolio demonstration of fine-grained multi-label emotion classification in English text. Not intended for mental-health assessment, hiring, insurance, moderation, or any decision about a person.

Training


Base model	`distilbert-base-uncased`
Data	GoEmotions `simplified` — 43,410 train / 5,426 val / 5,427 test
Objective	`BCEWithLogitsLoss` (28 independent sigmoid outputs)
Epochs	4
Batch size	32
Learning rate	3e-5, linear warmup (10%) then linear decay
Max sequence length	64 tokens
Other	gradient clipping at norm 1.0

Evaluation

Per-class decision thresholds were tuned on the validation split and applied frozen to the test split. Fill in your own numbers after training:

Metric	Flat 0.5 threshold	Tuned per-class
Micro F1	0.584	0.598
Macro F1	0.441	0.517

Read macro F1, not micro. Micro F1 is dominated by frequent labels; macro gives grief the same weight as neutral and is the honest measure of whether 28 distinct things were learned.

Inference

The thresholds are part of the model. Loading the weights without them will give you worse results than reported. They ship in inference_config.json alongside the taxonomy used by the demo Space.

python

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

MODEL_ID = "RenDeniz/emotion-detective-v2"
tok = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID).eval()
cfg = json.load(open(hf_hub_download(MODEL_ID, "inference_config.json")))

enc = tok("I can't keep up with any of this", return_tensors="pt",
          truncation=True, max_length=cfg["max_length"])
with torch.no_grad():
    probs = torch.sigmoid(model(**enc).logits).squeeze(0)

for i, p in enumerate(probs):
    name = model.config.id2label[i]
    if p >= cfg["thresholds"][name]:
        print(f"{name}: {p:.2f}")

Limitations and bias

Domain. Trained on Reddit comments annotated by a small pool of English speakers. It carries that platform's register and demographic skew, and degrades on other registers.
Long tail. grief, relief and pride have very few training examples and perform near zero. This is a data limitation, not a tuning one.
Sarcasm and negation are unreliable.
Emotion inference has a history of misuse in hiring, insurance pricing and attention monitoring. Don't use this for any of that.

License

Apache 2.0, matching the GoEmotions dataset.