te-sla/serbian-emotion-teslaxlm

Serbian multi-label emotion recognition — TeslaXLM

Model description

This is the primary publication encoder and operational default for Serbian social-media emotion recognition. It is a full fine-tune of te-sla/TeslaXLM at pinned base revision 982ff4401029ef0b0428e3704bbc592d9cf6965e.

The model predicts any valid subset of eight labels, in this fixed order: anger, anticipation, disgust, fear, joy, sadness, surprise, and trust. neutral is not an output label.

The release artifact is the surviving prespecified seed-42 Round 2 checkpoint. Its model.safetensors SHA-256 is d268f017dcd632d3895054d27c9cd34e7a5257ec8b695c1589ea45425cb11421. It uses the same cleaned cohort, pinned base revision, full fine-tuning, emotion-only hard labels, and unweighted BCE recipe as the final family. Its per-label thresholds were selected on validation only. It is a runnable representative; the separate ten-seed Round 2.1 panel remains the source of the model-family claim.

Intended use

Use the model for research on Serbian multi-label emotion classification and as the primary encoder in the accompanying experimental portfolio. It is suitable for batch analysis of Serbian Reddit/Twitter-style text after applying the same cleaning contract used during training.

It is not intended for clinical, employment, policing, moderation enforcement, or other high-impact decisions. It should not be used to infer a person's stable psychological state from a short post.

Data and preprocessing

Training used a cleaned, neutral-free Emo.SR Reddit and Twitter cohort with 29,961 rows: 23,961 train, 3,000 validation, and 3,000 test. Every retained row has one or two positive labels. Cleaning included Unicode normalization, Markdown-to-text conversion, whitespace control, URL/mention placeholders, hashtag-marker removal, punctuation-run capping, and emoji preservation.

The dataset is not distributed in this model repository. The model license metadata follows the pinned upstream base model; it does not grant rights to the training data. Dataset redistribution and privacy require a separate review.

Training

• Adaptation: full fine-tuning; no LoRA.
• Objective: emotion-only unweighted binary cross-entropy with hard labels.
• Maximum sequence length: 128 tokens.
• Learning rate: 2e-5.
• Effective batch size: 64 with physical-batch fallback.
• Precision: BF16 with TF32 and SDPA.
• Maximum epochs: 20; early-stopping patience: 3.
• Checkpoint rule: minimum validation loss within each fit.
• Release artifact: prespecified seed-42 Round 2 representative. This artifact was not chosen by test performance or substituted from a different objective.
• Decision thresholds: fitted per label on validation after checkpoint selection and stored in thresholds.json.

Weighted BCE, multitask objectives, GradNorm, and soft/hybrid targets were evaluated in later rounds but did not pass the frozen promotion gates. They are not part of this release.

Evaluation

The compatible ten-seed unweighted encoder panel used the same 3,000-item cleaned test cohort and validation-selected thresholds.

Scope	Seeds	Macro F1	Macro-F1 sample SD	Sample Jaccard
TeslaXLM recipe aggregate	10	0.558964	0.003614	0.560729
Released seed-42 artifact	1	0.558940	—	0.560611

The aggregate is the publication claim; it is not presented as the single artifact's score. TeslaXLM ranked first within this compatible replicated encoder panel and had the lowest Macro-F1 seed variation of the three final encoder families.

Per-label mean F1 across ten seeds was strongest for anger (0.7296), joy (0.6995), trust (0.6834), and sadness (0.5955). Disgust (0.3618) and fear (0.3588) remained the hardest labels.

The separate 2,575-row neutral abstention probe is a secondary diagnostic, not an F1 test set. The seed-42 family diagnostic activated at least one emotion on 95.88% of probe rows, so the model must not be described as a reliable neutral detector.

Inference

The published repository must include thresholds.json. Do not replace its values with 0.5 or tune them on evaluation data.

import json

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

repo_id = "te-sla/serbian-emotion-teslaxlm"
labels = [
    "anger", "anticipation", "disgust", "fear",
    "joy", "sadness", "surprise", "trust",
]

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id).eval()
with open(hf_hub_download(repo_id, "thresholds.json"), encoding="utf-8") as handle:
    thresholds = torch.tensor(json.load(handle)["thresholds"])

inputs = tokenizer("Ovo je primer teksta.", return_tensors="pt", truncation=True, max_length=128)
with torch.inference_mode():
    probabilities = torch.sigmoid(model(**inputs).logits[0]).cpu()
prediction = [label for label, score, cutoff in zip(labels, probabilities, thresholds) if score >= cutoff]
print(prediction)

Limitations

• The evidence is Serbian social-media text; generalization to news, speech, other languages, or formal prose is not established.
• Rare labels, especially disgust and fear, have substantially lower F1.
• The neutral probe shows high false activation under the selected threshold policy.
• Social-media models can reproduce demographic, topical, and annotation biases from their training data.
• Test evidence is internal and was reused adaptively across later rounds.
• Scores must not be pooled with incompatible nine-label, fixed-threshold, or generative panels.

Reproducibility and evidence

The technical report, exact configuration, ten-seed tables, threshold files, artifact hash, and code paths are indexed in the accompanying source repository. artifact_evaluation.json records this checkpoint's validation and test metrics, prediction-cache hashes, thresholds, and the separate family aggregate. The source repository is currently private; a public code and evidence archive is not yet linked from this card.

Citation

The final paper identifier and citation have not yet been assigned. The verified paper link and BibTeX citation will be added when available.