A 10-class emotion classifier fine-tuned for conversational AI memory systems. Uses an Inside Out 2 inspired taxonomy. v2 extends Kova-Mind/emotion-v1 with targeted training on pipeline-style bracket-tag fragments (X [positive], X [negative], X [neutral]) and entity fragments with embedded sentiment.

What v2 fixes

v1 shipped with excellent accuracy on natural conversational sentences, but two real-world failure modes appeared in production patterns:

Bracket-tag blindness. Fragments like hiking [positive] — Weekends activity. were consistently labeled neutral instead of joy. The [positive] / [negative] bracket is a Kova pipeline metadata tag the model had never seen in training.
Entity fragments with embedded sentiment. Fragments like lake (place) — Location where Rex loves swimming. were labeled neutral because v1 read only the nominal text; it missed the embedded positive signal.

v2 is trained on 706 additional Opus-labeled Kova-style examples that target these gaps.

Head-to-head vs v1

Evaluated against Claude Opus as fragment-only oracle (Opus reads each fragment in isolation and labels what it expresses).

Benchmark	n	v1	v2
21 production bracket-tag patterns	21	17/21 = 81.0 %	21/21 = 100.0 %
17 adversarial sanity cases	17	14/17 = 82.4 %	15/17 = 88.2 %
50 recent production patterns	50	42/50 = 84.0 %	40/50 = 80.0 %
Combined	88	73/88 = 83.0 %	76/88 = 86.4 %

+3.4 pts combined accuracy
+19 pts on bracket-tag patterns (the target gap)
Zero regressions on the 21 bracket patterns
The v2 dip on the 50-pattern set is 2 cases where v2 over-called joy on entity fragments with incidental third-party verbs; both v1 and v2 stayed within ±5 pts noise of the Opus oracle scoring.

Latency

Inference latency unchanged from v1: ~5 ms per call on an A40 GPU (after warmup). Well under the 10 ms production ceiling.

Classes

10-class Inside Out 2 taxonomy:

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

Same label IDs as v1 — drop-in replacement for any code loading Kova-Mind/emotion-v1.

Training Methodology

Same recipe as v1, with expanded domain data:

Base model: SamLowe/roberta-base-go_emotions
Head: fresh 10-class classification head (random init, 28→10 reinit from base)
Domain data (706 Opus-labeled examples, 3× weight):
- 21 real production bracket-tag patterns labeled by Opus
- 500 synthetic bracket-tag examples across all 10 emotions
- 185 natural conversational sentences (regression protection for general text)
Background data: Google GoEmotions simplified config, 28 labels remapped to the 10 Inside Out classes (rarest-label-wins tie-break)
Balancing: max 3000 samples/class, min 200 samples/class
Hyperparameters: 4 epochs, batch 32, learning rate 2e-5, AdamW, FP16, seed 42, warmup ratio 0.1, weight decay 0.01
Best checkpoint: epoch 2 (auto-selected via load_best_model_at_end on macro F1)
Hardware: 1× NVIDIA A40 (48GB)
Wall time: ~3 min 18 s pure training

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Kova-Mind/emotion-v2")
model = AutoModelForSequenceClassification.from_pretrained("Kova-Mind/emotion-v2")
model.eval()

text = "croissants [positive] — User A loves them exclusively."
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})")
# → joy (0.97)

Or with the pipeline API:

python

from transformers import pipeline

clf = pipeline("text-classification", model="Kova-Mind/emotion-v2", top_k=None)
print(clf("hiking [positive] — Weekends activity."))
# → joy (top class)

Limitations

English and a few non-English bracket fragments only. Non-English bracket patterns were included as training data (Arabic, Chinese) but broader multilingual coverage is out of scope.
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.
Single-label output. Trained with single-label resolution (rarest-wins tie-break). For multi-label emotion, a different head is needed.
Entity-fragment sentiment. v2 occasionally over-calls joy on factual third-party fragments with incidental positive verbs (e.g. "Alex (person) — Promoted to VP at Google"). v3 will target this with a curated third-party-factual training set.
envy class weakly supervised. GoEmotions does not contain an envy label; v2 sees only ~37 domain envy examples. Production coverage is limited.

Changelog from v1

+500 synthetic bracket-tag training examples (all 10 emotions)
+21 real production bracket-tag patterns (Opus-labeled)
+185 natural-sentence regression examples (Kova-style conversational text)
Same base model, same 10-class taxonomy, same hyperparameters, same label IDs.

Citation

bibtex

@misc{capo2026kovamindemotionv2,
  author       = {Capo, Alejandro},
  title        = {KovaMind Emotion v2: Bracket-Tag-Aware Emotion Classification for Conversational Memory Pipelines},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Kova-Mind/emotion-v2}}
}

Built By

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