This page lists the datasets, models and benchmarks released within the ERC Advanced Grant CLIMB (Cognitively-Inspired Language Models for Investigating Human Language Acquisition and Processing, Grant No. 101266749), as well as related resources developed by the CoLing Lab in recent years.
The resources are available on 🤗Hugging Face and 
GitHub.
Datasets
BAMBI
An Italian corpus of about 120 million tokens, organized into four developmental tiers, for training baby language models in a BabyLM-style setting.
Reference: Suozzi et al. (2025). BAMBI: Developing Baby Language Models for Italian. arXiv:2503.09481
CLASS-IT
A small-scale instruction-tuning dataset of about 130,000 examples. The conversational part comes from Switchboard and the educational part from Simple Wikipedia.
Reference: Capone, Bondielli and Lenci (2025). CLASS-IT: Conversational and Lecture-Aligned Small-Scale Instruction Tuning for BabyLMs. arXiv:2510.25364
Models
ConcreteGPT-124M
A 124M-parameter language model with the GPT-2 architecture and a 128-token context, trained in the BabyLM Strict-Small setting.
CLASS_IT
A 140M-parameter language model with the LLaMA architecture. It was pre-trained on the BabyLM strict-track data and then instruction-tuned on the CLASS-IT dataset.
Reference: Capone, Bondielli and Lenci (2025). arXiv:2510.25364
Benchmarks
CLEVER – Computational and Linguistic bEnchmarks for the study of VErb argument structuRe
A set of benchmarks for studying how language models handle verb argument structure.