QuantaSparkLabs/Undertone-GoEmotion

QuantaSparkLabs · Multi-label classifier

Undertone

A 28-label emotion tagger. One text, many colors.

distilbert · base go_emotions · simplified 28 labels apache-2.0 fast

Results · test split · tuned thresholds

Micro F1

0.584

+4.9 pts over 0.5 threshold

Macro F1

0.500

+2.5 pts over 0.5 threshold

Val → Test gap

0.006

micro · no overfit

The 28 emotions

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

Usage

from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="QuantaSparkLabs/Undertone-GoEmotion",
    top_k=None,
    function_to_apply="sigmoid",
)

text = "I can't believe this worked, thank you!"
scores = clf(text)[0]

# Apply per-label thresholds from thresholds.json
# instead of a blanket 0.5 for calibrated decisions.

Specification

Base

distilbert-base-uncased

Head

28-output linear, sigmoid, BCEWithLogitsLoss

Epochs

3 · lr 3e-5 · warmup 400 steps

Batch / len

32 train · 64 eval · max_len 128

Precision

fp16 mixed precision

Seed

42

Class weight

pos_weight, clipped [1, 10]

Thresholds

per-label grid over [0.10, 0.70], rounded to 0.02

Runtime

~5 min on a single consumer GPU

Limitations

01

Trained on Reddit comments. Domain shift on formal text — legal, medical, academic — is expected and measurable.

02

Truncated at 128 tokens. Long documents are only partially seen; outputs reflect the first chunk.

03

Rare labels — grief, pride, relief, remorse — have low test support. Check per-label F1 before depending on them.

04

Emotion annotations are subjective and culturally biased. The model inherits both.

05

Not a mental-health or diagnostic tool. Signals, not decisions. Human review required for any consequential use.

One text, many colors.

QuantaSparkLabs · 2026