Foundation models & labs for advanced learners
684 resources · page 22 of 22
Samrómur Milljón: An ASR Corpus of One Million Verified Read Prompts in Icelandic
by Carlos Daniel Hernández Mena et al. (Reykjavik University Language and Voice Lab)
LREC-COLING 2024 paper presenting Samrómur Milljón, an Icelandic ASR corpus of one million validated read prompts.
Sakana blog
by Sakana AI
Sakana AI blog with research posts on evolutionary model merging, agents and its Fugu models, partly in Japanese.
ReLiK: Retrieve and LinK, Fast and Accurate Entity Linking and Relation Extraction on an Academic Budget
by Riccardo Orlando et al. (Sapienza University of Rome)
arXiv paper (2024) introducing ReLiK, fast entity linking and relation extraction by Sapienza NLP.
SEQUOR: A Multi-Turn Benchmark for Realistic Constraint Following
by Beatriz Canaverde et al. (Instituto de Telecomunicações)
arXiv paper (2026) introducing SEQUOR, a multi-turn benchmark for realistic constraint following.
SmolLM3 model card
by Hugging Face Smol Models Research
Hugging Face model card for SmolLM3-3B with architecture, training details and evaluation results.
Bielik v3 Small: Technical Report
by Krzysztof Ociepa et al. (SpeakLeash)
arXiv paper (2025) Bielik v3 Small technical report on SpeakLeash's small Polish language models.
Apertus: Democratizing Open and Compliant LLMs for Global Language Environments
by Project Apertus (Swiss AI Initiative)
arXiv paper (2025) introducing Apertus, the Swiss AI Initiative's open and compliant multilingual LLMs.
T-Pro 2.1 instruction-following and tool-calling report
by T-Bank
T-Bank Habr article in Russian on T-Pro 2.1 improvements to instruction following and tool calling.
Connectionism (Thinking Machines Lab)
Thinking Machines Lab blog Connectionism, with research posts on training, numerics and infrastructure experiments.
Pretraining Finnish ModernBERTs
by Akseli Reunamo et al. (University of Turku)
arXiv paper (2025) by TurkuNLP on pretraining Finnish ModernBERT encoder models.
UK AISI work
UK AI Security Institute work page listing its research, evaluations and blog posts.
Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian
by Aleksandr Nikolich et al. (ITMO University)
arXiv paper (2024) introducing Vikhr, a family of open instruction-tuned Russian language models.