Kenji-Endo: a BabyLM @EVALITA

Calogero Jerik Scozzaro; Mattero Rinaldi; Gianluca Mittone; M.A. Stranisci

Kenji-Endo: a BabyLM @EVALITA

Calogero Jerik Scozzaro
, Mattero Rinaldi
, Gianluca Mittone
, M.A. Stranisci

University of Torino
Aequa-tech

Research output: Conference Article in Proceeding or Book/Report chapter › Article in proceedings › Research › peer-review

Abstract

We present Kenji-Endo, a BabyLM pretrained on a dedicated Italian dataset, which participated in four tasks at the 9th edition of EVALITA: DeSegMa, MultiPRIDE, IMPOLS, and FadeIT. Kenji-Endo achieved competitive performance across all tasks, demonstrating that language modeling with limited data and compact model sizes
can represent a viable alternative to Large Language Models

Original language	Undefined/Unknown
Title of host publication	Ceur Workshop Proceedings
Number of pages	11
Publisher	CEUR Workshop Proceedings
Publication date	2026
Pages	16-26
Publication status	Published - 2026
Externally published	Yes
Event	9th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop - Bari, Italy Duration: 26 Feb 2026 → 27 Feb 2026 Conference number: 9 https://www.evalita.it/campaigns/evalita-2026/

Conference

Conference	9th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop
Number	9
Country/Territory	Italy
City	Bari
Period	26/02/2026 → 27/02/2026
Internet address	https://www.evalita.it/campaigns/evalita-2026/

Keywords

NLP
Evaluation
Italian
Language Models
BabyLM

Access to Document

https://ceur-ws.org/Vol-4195/36.pdfLicence: CC BY

Cite this

@inproceedings{4483e4ab9c4b47089637d9e09107cfdc,

title = "Kenji-Endo: a BabyLM @EVALITA",

abstract = "We present Kenji-Endo, a BabyLM pretrained on a dedicated Italian dataset, which participated in four tasks at the 9th edition of EVALITA: DeSegMa, MultiPRIDE, IMPOLS, and FadeIT. Kenji-Endo achieved competitive performance across all tasks, demonstrating that language modeling with limited data and compact model sizes can represent a viable alternative to Large Language Models",

keywords = "NLP, Evaluation, Italian, Language Models, BabyLM",

author = "Scozzaro, \{Calogero Jerik\} and Mattero Rinaldi and Gianluca Mittone and M.A. Stranisci",

year = "2026",

language = "Udefineret/Ukendt",

pages = "16--26",

booktitle = "Ceur Workshop Proceedings",

publisher = "CEUR Workshop Proceedings",

note = "9th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop ; Conference date: 26-02-2026 Through 27-02-2026",

url = "https://www.evalita.it/campaigns/evalita-2026/",

}

TY - GEN

T1 - Kenji-Endo: a BabyLM @EVALITA

AU - Scozzaro, Calogero Jerik

AU - Rinaldi, Mattero

AU - Mittone, Gianluca

AU - Stranisci, M.A.

N1 - Conference code: 9

PY - 2026

Y1 - 2026

N2 - We present Kenji-Endo, a BabyLM pretrained on a dedicated Italian dataset, which participated in four tasks at the 9th edition of EVALITA: DeSegMa, MultiPRIDE, IMPOLS, and FadeIT. Kenji-Endo achieved competitive performance across all tasks, demonstrating that language modeling with limited data and compact model sizes can represent a viable alternative to Large Language Models

AB - We present Kenji-Endo, a BabyLM pretrained on a dedicated Italian dataset, which participated in four tasks at the 9th edition of EVALITA: DeSegMa, MultiPRIDE, IMPOLS, and FadeIT. Kenji-Endo achieved competitive performance across all tasks, demonstrating that language modeling with limited data and compact model sizes can represent a viable alternative to Large Language Models

KW - NLP

KW - Evaluation

KW - Italian

KW - Language Models

KW - BabyLM

M3 - Konferencebidrag i proceedings

SP - 16

EP - 26

BT - Ceur Workshop Proceedings

PB - CEUR Workshop Proceedings

T2 - 9th Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop

Y2 - 26 February 2026 through 27 February 2026

ER -