Kova-Mind/emotion-v1

KovaMind Emotion v1

A 10-class emotion classifier fine-tuned for conversational AI memory systems. Uses an Inside Out 2 inspired taxonomy optimized for the Kova Mind agent memory pipeline.

Performance

Head-to-head on a 53-sentence curated evaluation set covering 9 of the 10 classes (A40 GPU, FP32):

Metric	Baseline (SamLowe/roberta-base-go_emotions, 28→10 mapped)	KovaMind Emotion v1
Accuracy	73.6% (39/53)	96.2% (51/53)
Avg inference latency	42 ms	7 ms

+22.6 percentage points and ~6× faster inference — from direct 10-class head + domain-adapted weights, without a post-hoc label mapping.

Per-class accuracy (KovaMind Emotion v1)

Class	Correct	Accuracy
joy	10/10	100%
sadness	8/8	100%
anger	6/7	85.7%
fear	6/6	100%
disgust	5/5	100%
anxiety	5/5	100%
embarrassment	3/4	75.0%
ennui	3/3	100%
neutral	5/5	100%

(The envy bucket was not exercised by this eval set — production monitoring tracks it separately.)

Classes

The 10 emotion buckets:

ID	Label
0	joy
1	sadness
2	anger
3	fear
4	disgust
5	anxiety
6	embarrassment
7	envy
8	ennui
9	neutral

This taxonomy is inspired by Pixar's Inside Out 2 — extending the original five (joy, sadness, anger, fear, disgust) with anxiety, embarrassment, envy, and ennui, plus a neutral bucket. It was chosen because it maps cleanly onto the emotions that drive behavioral memory encoding in conversational contexts.

Training Methodology

• Base model: SamLowe/roberta-base-go_emotions — RoBERTa-base already pre-adapted to emotion classification on GoEmotions (28 labels)
• Head: fresh 10-class classification head (random init)
• Domain data: 433 Opus-labeled conversational sentences, targeted at the weak buckets (anxiety, embarrassment, envy, ennui) — weighted 3x during training for high signal
• Background data: Google GoEmotions simplified config, 28 labels remapped to the 10 Inside Out classes (priority order: rarest-label-wins for multi-label cases)
• Balancing: max 3000 samples/class, min 200 samples/class
• Hyperparameters: 4 epochs, batch 32, learning rate 2e-5, AdamW, FP32
• Hardware: 1× NVIDIA A40 (48GB)
• Wall time: ~12 minutes

Why Domain-Specific Beats Larger Models

Prior to this fine-tune we benchmarked microsoft/deberta-v3-large fine-tuned on GoEmotions. It showed no meaningful gain over the RoBERTa-base baseline for our 10-class downstream task. The real win came from three choices:

1. Conversational training data — 433 Opus-labeled sentences from real agent-memory contexts, not Reddit comments.
2. Reduced label space — 10 clear buckets vs GoEmotions' 28 (many of which are sparsely populated or overlap).
3. High-agreement labels — Opus-labeled set had strong inter-annotator consistency on the target labels.

Larger model ≠ better when the bottleneck is label quality + domain match.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Kova-Mind/emotion-v1")
model = AutoModelForSequenceClassification.from_pretrained("Kova-Mind/emotion-v1")
model.eval()

text = "I just got the job!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=256)
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)
    pred_id = torch.argmax(probs, dim=-1).item()

print(f"{model.config.id2label[pred_id]} ({probs[0, pred_id]:.3f})")

Or with the pipeline API:

from transformers import pipeline

clf = pipeline("text-classification", model="Kova-Mind/emotion-v1", top_k=None)
print(clf("I have a huge presentation tomorrow and I can't stop overthinking it."))
# → anxiety (top class)

Limitations

• English only. No multilingual training data.
• Conversational AI context. Optimized for agent-memory ingestion of natural user dialog. May underperform on social-media, legal, or heavily-stylized text.
• 10-class granularity. Does not capture finer emotional shades (e.g. admiration vs pride, grief vs sadness, amusement vs joy). Those get absorbed into the closest bucket.
• Opinionated taxonomy. The Inside Out 2 bias toward anxiety / embarrassment / envy / ennui reflects what matters for long-term memory encoding in agents; other applications may want a different split.
• Single-label output. Trained with single-label resolution (rarest-wins tie-break). For multi-label emotion, a different head is needed.

Citation

@misc{capo2026kovamindemotion,
  author       = {Capo, Alejandro},
  title        = {KovaMind Emotion v1: Domain-Specific Fine-Tuning Beats Larger Models for Conversational Emotion Classification},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Kova-Mind/emotion-v1}}
}

Built By

kovamind.io — private, customer-owned AI memory infrastructure.