AI

We present RoBaseLM-S (125 M) and RoBaseLM-M (260 M), two compact Romanian decoder-only language models trained from scratch on a 4.3 B-token curated corpus. Architecturally, they follow a modern LLaMA-style recipe with pre-norm RMSNorm, rotary position embeddings, SwiGLU feed-forward blocks, grouped-query attention, and 4 k-token context windows. We release both full-precision (FP16) and post-training 5-bit (Q5_K_M) checkpoints in GGUF format for lightweight local inference. The 5-bit variants fit under 500 MB and generate text in real time on a Jetson Nano 4 GB, enabling fully offline Romanian text generation on consumer-grade edge hardware. We evaluate the models intrinsically (multi-domain perplexity across news, literary prose, poetry, and heterogeneous web text) and extrinsically (LaRoSeDa sentiment classification and RO-STS sentence similarity). Relative to Romanian GPT-2–style baselines at similar parameter scales, RoBaseLM-S and RoBaseLM-M reduce perplexity substantially, e.g., from 30.7 to 15.9 on our held-out news split. The 5-bit post-training quantized checkpoints remain within FP16 performance across all reported tasks. To our knowledge, these are the first Romanian small language models explicitly optimized for long-context inference, post-training quantization, and low-power on-device deployment.

Anticoagulant pharmacology is a cognitively demanding domain in undergraduate medical education, with persistent challenges in learner engagement, retention, and safe clinical application. Cinematic Clinical Narratives (CCNs) offer a theory-informed multimedia approach designed to support learning through narrative structure, visual mnemonics, and affective engagement. We conducted a multi-site quasi-experimental study within a six-week Cancer, Hormones, and Blood course across a distributed medical education program. First-year medical students received either a traditional case-based lecture or an animated CCN (Twilight: Breaking Clots) during a one-hour anticoagulant pharmacology session. Learning outcomes were assessed using pre- and posttests, learner engagement was measured with the Situational Interest Survey for Multimedia (SIS-M), and exploratory eye tracking with second-year medical students was used to assess visual attention to embedded mnemonics. Both instructional groups demonstrated significant learning gains, with fold-change analyses indicating greater relative improvement among students exposed to the CCN. The animated CCN elicited significantly higher triggered situational interest compared with non-animated cases (p = 0.019), while also being preferred by the majority of participants. Qualitative analysis revealed that learners perceived CCNs as particularly effective for initial encoding and memorization, while non-animated cases supported subsequent clinical application. Eye-tracking data demonstrated high visual uptake and sustained attention to key mnemonic elements. Together, these findings support expert-designed, genAI-assisted CCNs as a validated and complementary instructional approach in medical education.

The categorization of real-time defects in heterogeneous domains is a long-standing challenge in the field of industrial visual inspection systems, primarily due to significant visual variations and the lack of labelled information in real-world inspection settings. This work presents the Adaptive Attention DropBlock (AADB) framework, a lightweight deep learning framework that was developed to promote cross-domain defect detection using attention-guided regularization. The proposed architecture integrates the Convolutional Block Attention Module (CBAM) and an organized DropBlock-based regularization scheme, creating a unified and robust framework. Although CBAM-based approaches improve localization of defect-related areas and traditional DropBlock provides a generic spatial regularization, neither of them alone is specifically designed to reduce domain overfitting. To address this limitation, AADB combines attention-directed feature refinement with a progressive, transfer-aware dropout policy that promotes the learning of domain-invariant representations. The proposed model is built on a MobileNetV2 base and trained through a two-phase transfer learning regime, where the first phase consists of pretraining on a source domain and the second phase consists of adaptation to a visually dissimilar target domain with constrained supervision. The overall analysis of a metal surface defect dataset (source domain) and an aircraft surface defect dataset (target domain) shows that AADB outperforms CBAM-only, DropBlock-only, and conventional MobileNetV2 models, with an overall accuracy of 91.06%, a macro-F1 of 0.912, and a Cohen’s k of 0.866. Improved feature separability and localization of error are further described by qualitative analyses using Principal Component Analysis (PCA) and Grad-CAM. Overall, the framework provides a practical, interpretable, and edge-deployable solution to the classification of cross-domain defects in the industrial inspection setting.