Language models and datasets

Code, datasets and fine-tuned language models released with SAILab's papers on natural language processing. Each release has a GitHub repository and a Hugging Face collection, and cites the paper it comes from. Most of them cover languages and tasks that large language models still handle poorly: Farsi, Arabic varieties, Sindhi, Turkish and Italian, educational content, and specialised domains such as pharmaceutics and industry.

The work is led by Kamyar Zeinalipour with Marco Gori, Marco Maggini, Monica Bianchini and colleagues in Siena and abroad. See also the language research line and the lab's news.

Benchmarks and data for languages with little data

• FarSense

  Farsi commonsense benchmark, 6 fine-tuned adapters.

  Paper: FarSense: A Comprehensive Commonsense Benchmark and Evaluation Framework for the Farsi Language (IJCNLP-AACL 2025).

  Code on GitHub Models and data on Hugging Face

• ArabicSense

  Arabic commonsense benchmark and models.

  Paper: ArabicSense: A Benchmark for Evaluating Commonsense Reasoning in Arabic with Large Language Models (WACL-4 2025).

  Code on GitHub Models and data on Hugging Face

• Shawarma Chats

  Dialogue benchmark in three Arabic varieties.

  Paper: Shawarma Chats: A Benchmark Dialogue and Evaluation Platter in Egyptian, Maghrebi and Modern Standard Arabic (ArabicNLP 2025).

  Code on GitHub Models and data on Hugging Face

• TalkFa

  Farsi dialogue generation and understanding.

  Paper: TalkFa: A Unified Benchmark for Farsi Dialogue Generation and Understanding (Preprint, arXiv:2609.01810).

  Code on GitHub Models and data on Hugging Face

• Sindhi Reasoning

  Culturally grounded instruction tuning for Sindhi.

  Paper: Enabling Structured Reasoning in Sindhi with Culturally Grounded Instruction Tuning (LoResLM 2026).

  Code on GitHub Models and data on Hugging Face

• PersianMCQ-Instruct

  Persian multiple-choice question generation.

  Paper: PersianMCQ-Instruct: A Comprehensive Resource for Generating Multiple-Choice Questions in Persian (LoResLM 2025).

  Code on GitHub Models and data on Hugging Face

Educational content

• MultiHint-9K

  Pedagogical hints, 9,000 of them, with five models.

  Paper: MultiHint-9K: Orchestrating Human and AI Agency for Scalable Pedagogical Hint Generation (AIED 2026 workshops, CEUR 4248).

  Code on GitHub Models and data on Hugging Face

• Italian crosswords

  Italian clue generation across three papers.

  Papers: Italian Crossword Generator: An In-depth Linguistic Analysis in Educational Word Puzzles (Italian Journal of Computational Linguistics 11); Harnessing LLMs for Educational Content-Driven Italian Crossword Generation (CLiC-it 2024); Italian Crossword Generator: Enhancing Education through Interactive Word Puzzles (CLiC-it 2023).

  Code on GitHub Models and data on Hugging Face

• Arabic crosswords

  Arabic clue generation, annotation interface.

  Papers: From Arabic Text to Puzzles: LLM-Driven Development of Arabic Educational Crosswords (LoResLM 2025); ArabIcros: AI-Powered Arabic Crossword Puzzle Generation for Educational Applications (ArabicNLP @ ACL 2023).

  Code on GitHub Models and data on Hugging Face

• Turkish crosswords

  Turkish clue generation models.

  Paper: A Turkish Educational Crossword Puzzle Generator (AIED 2024).

  Code on GitHub Models and data on Hugging Face

• Turkish quizzes

  Quiz generation from Turkish educational text.

  Paper: Automating Turkish Educational Quiz Generation Using Large Language Models (ISPR 2024).

  Code on GitHub Models and data on Hugging Face

• Essay syntax feedback

  Automated syntax feedback for student writing.

  Paper: Advancing Student Writing Through Automated Syntax Feedback (AIEER 2024).

  Code on GitHub Models and data on Hugging Face

Science and industry

• Round-trip KG

  Text to knowledge graph and back, with fidelity scores.

  Paper: From Graph to Text and Back: Semantic Fidelity in Automated Industrial Knowledge Graphs (ACL 2026, industry track).

  Code on GitHub Models and data on Hugging Face

• PharmaQA.IT

  Italian question answering over regulatory documents.

  Paper: PharmaQA.IT: an Italian dataset for Q&A in the pharmaceutical domain (EACL 2026, industry track).

  Code on GitHub Models and data on Hugging Face

• PharMistral

  Code and models for the CMPB Update 2026 paper.

  Paper: PharMistral: Drug generation for novel protein targets by Mistral Large Language Model (Computer Methods and Programs in Biomedicine Update 10).

  Code on GitHub Models and data on Hugging Face

• Protein design

  Protein sequence generation with language models.

  Paper: Design Proteins Using Large Language Models: Enhancements and Comparative Analyses (Language + Molecules @ ACL 2024).

  Code on GitHub Models and data on Hugging Face

Language and cognition

• Crossword Surprisal

  Clue difficulty predicted from token probabilities.

  Paper: Surprisal and Crossword Clues Difficulty: Evaluating Linguistic Processing between LLMs and Humans (CLiC-it 2025).

  Code on GitHub Models and data on Hugging Face

Open European language models

Kamyar Zeinalipour also works on open language models for Europe's languages at Villanova.AI. He built Villanova-SFT-2603, the instruction-tuning data for the first-generation Villanova-2B models (1,711,114 conversations in five languages, Apache 2.0). The second generation, Villanova2, released in November 2026, cites the paper Multi-SFT: Permissively-Licensed Multilingual Instruction-Following.