VSR Seminar with Two Talks

1st talk: Topological transport in real materials from ab initio

Speaker:
Theodoros Adamantopoulos, Peter Grünberg Institut, PGI-1

Abstract:

The research conducted within our project has served as a foundation for several directions in computational spintronics in the past, among which one of the most recent and promising is orbitronics. As a field, orbitronics relies on the idea that the orbital degree of freedom of electrons, complementary to spin-orbit-driven spintronic effects, offers an efficient channel for angular momentum transport and magnetization dynamics. Our project uses highly-predictive ab initio electronic-structure methods, built on density-functional theory combined with Wannier-function interpolation and Berry-curvature response theory, implemented in massively parallelized codes, to compute and predict various properties of orbital currents, orbital torques, and orbital angular momentum dynamics in real materials at the atomistic level. Within the course of the project, in the recent past, we discovered a key orbitronics effect of orbital pumping — the generation of orbital angular momentum currents by magnetization dynamics, established how crystal structure governs the propagation of orbital currents and the torques they exert on magnetization, and predicted a giant orbital magnetoresistance effect in an antiferromagnetic insulator, with further orbital effects identified in altermagnets and in two-dimensional dichalcogenides. These results relied on large-scale, highly parallel Wannier-interpolation-based first-principles calculations whose favorable scaling made efficient use of the allocated computing time. The talk will summarize the key scientific findings from recent past, methods, codes, and parallelization strategies underlying this Wannier-based orbitronics framework, briefly quantifying computing resources consumed so far and provide an estimate for the resource needs anticipated for the coming period.

2nd Talk: Synthlaion: Synthetic data generation and dataset composition for pre- and post-training of generalist open foundation models with strong reasoning and agentic function

Speaker:
Jenia Jitsev, Jülich Supercomputing Centre, JSC

Abstract:

Strong open foundation models increasingly depend on synthetic data: reasoning traces, agentic traces and rephrased web text generated by strong teacher models. In the VSR project synthlaion on JURECA we develop open, reproducible recipes for generating and composing such data. The talk presents works we performed in this direction: i) OpenThoughts (ICLR 2026 Oral), where over 1000 controlled ablations yield a dataset of 1.2M reasoning traces and 7B to 32B models that matched the strongest open reasoning models, including the finding that the strongest model is not the best generator; ii) OpenThoughts-Agent, which carries the same recipe search to agentic traces and Terminal-Bench style tasks; and iii) MixtureVitae (TMLR 2026, J2C certification), a permissive-first pre-training corpus in which front-loaded synthetic reasoning and instruction data gives a 1.7B base model math and code skills otherwise reached only after post-training. We describe the codes and parallelization used for distributed generation (vLLM, SGLang), supervised fine-tuning (LlamaFactory) and reinforcement learning with executable environments (SkyRL, Harbor), their measured scaling, and the resources needed to scale generation, teacher comparison and post-training.