We describe our submission to Task 1 of #SMM4H-HeaRD 2026: multilingual binary classification of adverse drug event (ADE) mentions in social media. Our system fine-tunes xlm-roberta-large with LoRA adapters and learned language embeddings, using two-stage training (CADEC translated domain adaptation, then five-fold cross-validation on the official training set). We ensemble the five fold checkpoints by mean logits, apply temperature scaling on the development set, and tune decision thresholds to maximize the official metric. On development, the final ensemble reaches macro-F₁ 0.788 with a global threshold and 0.796 with per-language thresholds; our best official test submission achieves macro-F₁ 0.616 (ID 678990).

The SemEval-2026 Task 3 is a Dimensional aspect-based sentiment analysis (DimABSA) task which extends traditional ABSA by predicting continuous regression in two dimensions: valence (V) and arousal (A). The Track A/Subtask 1 represent multilingual task in which for a given text and aspects mentioned in it, there is a need to predict V/A scores for each aspect. Our approach is based on the pretraining-finetuning concept: we first pretrain multilingual model (M ′) followed by its fine-tuning (M ′′ l,d) on the training data of specific domain (d) and language (l). We adopt XLM-RoBERTa (M ) as the encoder with separate regression heads for valence and arousal prediction. Our experiments on manual split of official SemEval-2026 Task 3 dataset (D20% train) demonstrate that fine-tuning model in two stages (M ′′ l,d) results in average ≈ 1.36 times improvement by RMSEva over approach of direct fine-tuning (Ml,d). To investigate limitations of the existing approach we visualize and discuss limitations of our system. Our code is publicly available.