albertmartinez/xlm-roberta-large-sdg-openalex-multilabel-2026-09

Xlm Roberta Large Sdg Openalex Multilabel 2026 09

Model description

Multilingual sequence classification model fine-tuned from XLM-RoBERTa-large to predict UN Sustainable Development Goal (SDG) labels (sdg1 … sdg17) from scientific text (OpenAlex-derived corpus).

• —Task: multi-label text classification
• —Classes: 17 SDG labels (sdg1–sdg17, incl. SDG 17)
• —Fine-tuned from: FacebookAI/xlm-roberta-large
• —Selection metric: F1 macro (best checkpoint)

Intended uses & limitations

Intended uses

• —Tagging research abstracts and publications with one or more SDG labels
• —Large-scale bibliometric SDG analytics on OpenAlex-style corpora
• —Multilingual SDG pipelines that require SDG 17 (Partnerships)

Out-of-scope uses

• —Legal or compliance decisions without human review
• —High-stakes allocation of funds based solely on model predictions
• —Fine-grained policy classification outside the SDG taxonomy

Limitations

• —Trained on sdg-openalex; performance may drop on non-academic or very short text
• —Input format: <TITLE>…</TITLE> and optionally <ABSTRACT>…</ABSTRACT> (same as the dataset text column; ~29% of training samples are title-only)
• —Strong class imbalance across SDGs; rare goals may underperform
• —Dataset majoritàriament single-label per fila (~0.0014% train amb >1 SDG); el model prediu multi-label via sigmoid per classe
• —Evaluation used the dataset test split as validation during training
• —Threshold 0.5 per defecte a l'exemple d'ús; cal ajustar segons el cas d'ús

Training and evaluation data

Dataset

• —Hub: albertmartinez/sdg-openalex
• —Config: v1
• —Train split: 1,041,829 examples
• —Test split: 260,458 examples (used for evaluation each epoch)
• —Input column: text
• —Label column: labels (sdg1 … sdg17)

Preprocessing

• —Tokenizer: XLM-RoBERTa (max_seq_length=512)
• —Pos weights applied to BCEWithLogitsLoss (inverse frequency, normalized)
• —Train shuffle: enabled

Training procedure

Framework

Fine-tuning with Hugging Face Transformers Trainer, using the official `run_classification.py` script (project fork with class-weight / pos_weight support).

Training

• —Optimizer: AdamW (Trainer default)
• —Precision: bf16 on NVIDIA GPU
• —Best checkpoint selected by eval F1 macro (load_best_model_at_end=True)
• —Early stopping patience: 1 (eval F1 macro)

Training hyperparameters

Training results

Final metrics (best checkpoint)

Eval per epoch

Per-class evaluation

Per-class metrics via confusion matrix no disponibles en multi-label; veure F1 macro global al bloc d'avaluació.

Confusion matrix

Per-class metrics via confusion matrix no disponibles en multi-label; veure F1 macro global al bloc d'avaluació.

Usage

El model s'ha entrenat amb el format del dataset sdg-openalex: <TITLE>…</TITLE><ABSTRACT>…</ABSTRACT>. Si només tens títol, <TITLE>…</TITLE> n'hi ha prou.

from transformers import pipeline

model_id = "albertmartinez/xlm-roberta-large-sdg-openalex-multilabel-2026-09"


def format_input(title: str, abstract: str | None = None) -> str:
    """Format d'entrada del dataset sdg-openalex."""
    text = "<TITLE>" + title.strip() + "</TITLE>"
    if abstract and abstract.strip():
        text += "<ABSTRACT>" + abstract.strip() + "</ABSTRACT>"
    return text


classifier = pipeline(
    "text-classification",
    model=model_id,
    top_k=None,  # totes les etiquetes (multi-label, sigmoid automàtic)
)

text = format_input(
    "Partnerships for sustainable development goals and international cooperation.",
    "This study examines cross-sector collaboration for SDG implementation.",
)
results = classifier(text)[0]

threshold = 0.5
for item in results:
    if item["score"] >= threshold:
        print(item["label"], item["score"])

Framework versions