boffire/kabyle-emotion-xlmr

Kabyle Emotion Classifier

A fine-tuned XLM-RoBERTa model for 7-class emotion recognition in Kabyle (Taqbaylit), a low-resource Afro-Asiatic and Amazigh language spoken in Algeria.

Model Details

Attribute	Value
Base model	xlm-roberta-base (fine-tuned from boffire/kabyle-emotion-xlmr)
Architecture	XLM-RoBERTa for Sequence Classification
Parameters	~278 M
Languages	Kabyle (kab)
Task	Text Classification (Emotion Detection)
Classes	7 — anger, disgust, fear, joy, sadness, surprise, neutral

Training Data

The model was trained via cross-lingual label transfer from English to Kabyle using parallel sentence pairs:

1. Round-trip parallel corpus (eng_kab_roundtrip_good.tsv) — 131,301 English–Kabyle sentence pairs with back-translation quality scores.
2. Tatoeba parallel corpus — 138,353 additional English–Kabyle linked sentences downloaded from tatoeba.org.

Labeling pipeline:

• English sentences were labeled with j-hartmann/emotion-english-distilroberta-base.
• Labels were transferred to the Kabyle side via sentence alignment.
• Low-confidence predictions (< 0.75) were filtered out.
• The neutral class was capped at 2,000 examples to reduce imbalance.

Final balanced dataset:

• Total labeled rows (raw): 225,036
• Final training set: 54,486 rows
  • joy: 12,539 | disgust: 11,983 | sadness: 9,666 | surprise: 6,418 | fear: 6,334 | anger: 5,546 | neutral: 2,000
• Train / Val / Test split: 40,864 / 5,449 / 8,173

Performance

Test Set Results (8,173 samples)

Emotion	Precision	Recall	F1-Score	Support
anger	0.70	0.75	0.73	832
disgust	0.81	0.64	0.72	1,797
fear	0.72	0.74	0.73	950
joy	0.82	0.80	0.81	1,881
sadness	0.72	0.77	0.75	1,450
surprise	0.87	0.87	0.87	963
neutral	0.22	0.39	0.28	300

• Accuracy: 0.74
• Weighted Avg F1: 0.75
• Macro Avg F1: 0.69

How to Use

Quick inference with transformers

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="boffire/kabyle-emotion-xlmr",
    device=0  # use -1 for CPU
)

# Example sentences
examples = [
    "Tafyirt 1",           # → sadness (0.98)
    "Tafyirt tis 2",       # → joy (0.72)
    "Tafyirt tis 3",    # → neutral / sadness
]

for text in examples:
    result = classifier(text, top_k=None)
    top = sorted(result[0], key=lambda x: x["score"], reverse=True)[0]
    print(f"{text} → {top['label']} ({top['score']:.3f})")

Loading the model directly

from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("boffire/kabyle-emotion-xlmr")
model = AutoModelForSequenceClassification.from_pretrained("boffire/kabyle-emotion-xlmr")

# Tokenize and predict
inputs = tokenizer("Imir-a, Muiriel tesɛa 20 n yiseggasen.", return_tensors="pt", truncation=True)
outputs = model(**inputs)

Training Details

Hyperparameter	Value
Epochs	6 (with early stopping, patience=2)
Batch size	32 per device (effective 64 with gradient accumulation)
Learning rate	2e-5
Max sequence length	96
Weight decay	0.01
Warmup steps	~10% of total steps
Mixed precision	FP16
Class weights	Balanced (sklearn.utils.class_weight.compute_class_weight)
Optimizer	AdamW (Hugging Face default)
Best checkpoint	Epoch 6 (loaded automatically via load_best_model_at_end)

Limitations & Caveats

1. Silver labels: Ground-truth emotions were projected from an English classifier. Some labels may not perfectly capture Kabyle cultural or emotional nuance.
2. Neutral class weakness: The neutral class performs poorly (F1 ~0.28) because it contains many low-confidence English predictions. Consider treating it as a "no strong emotion" fallback rather than a reliable label.
3. Translation quality: The parallel corpus includes round-trip translated sentences. Imperfect translations may introduce label noise.
4. No native speaker validation: The test set was held out from the same silver-labeled pool. A small native-annotated benchmark would give a more accurate human ceiling.
5. Imbalanced source: Tatoeba data is skewed toward simple, short sentences. Performance may degrade on longer, more complex Kabyle text (social media, literature, etc.).

Intended Use

• Research in low-resource NLP and Afro-Asiatic language processing.
• Downstream applications requiring coarse emotion signals in Kabyle text (e.g., content moderation, mental-health screening, customer feedback analysis).
• Baseline for future Kabyle emotion models trained on native annotations.

Citation

If you use this model, please cite:

@misc{boffire_kabyle_emotion_xlmr,
  title = {Kabyle Emotion Classifier},
  author = {Boffire},
  year = {2026},
  howpublished = {\url{https://huggingface.co/boffire/kabyle-emotion-xlmr}},
  note = {Fine-tuned XLM-RoBERTa for 7-class emotion detection in Kabyle via cross-lingual label transfer from English}
}

License

This model is released under the Apache 2.0 license. The base XLM-RoBERTa weights and the English emotion classifier (j-hartmann/emotion-english-distilroberta-base) are subject to their respective original licenses.

Acknowledgments

• Tatoeba Project for the English–Kabyle parallel corpus.
• j-hartmann for the English emotion classifier used for label projection.
• Hugging Face transformers, datasets, and accelerate teams for the training infrastructure.