colinglab/BAMBI_dataset

🦌 BAMBI Dataset

Dataset Summary

The BAMBI Dataset is an Italian-language corpus designed to provide ecologically valid and cognitively plausible linguistic input for BabyLM-style language model training. It is organized into four developmental tiers, corresponding to different phases of linguistic growth in childhood and adolescence:

Tier	Age Range (years)	Approx. Tokens	Description
Tier 1	0–6	~26M	Early childhood and preschool exposure
Tier 2	6–12	~31M	Primary school and literacy development
Tier 3	12–18	~32M	Adolescent linguistic expansion
Tier 4	18–24	~31M	Young adult and advanced language exposure

Each tier reflects age-appropriate linguistic input, curated from authentic Italian sources for ecological validity and developmental plausibility.

---

Data Composition

The corpus integrates transcriptions of oral materials that mirror the linguistic environment of Italian speakers from early childhood through young adulthood.

Source Types

• Entertainment: audiobooks, TV shows, podcasts, and films
• Educational: textbooks, didactic materials, and learning media
• News and informational: simplified news, magazines, and public media
• Caregiver and spontaneous speech: naturalistic speech and conversational transcriptions

Each document includes metadata indicating:

• Source type
• Data origin
• Intended age tier

---

Linguistic Characteristics

• Language: Italian 🇮🇹
• Register: naturalistic, age-appropriate, and cognitively plausible speech
• Total tokens: ~120 million
• Format: plain text (.csv) with detailed metadata

---

Intended Use

The BAMBI dataset is intended for:

• Training BabyLM-style or other small-scale developmental language models in Italian
• Research on cognitive and linguistic development
• Curriculum learning and ecological modeling of language acquisition

Not intended for commercial NLP applications.

---

Ethical Considerations

TBD

Citation

TBD

More about BAMBI

Suozzi, A., Capone, L., Lebani, G. E., & Lenci, A. (2025). BAMBI: Developing Baby Language Models for Italian. arXiv preprint arXiv:2503.09481.

Capone, L., Suozzi, A., Lebani, G. E., & Lenci, A. (2025). BAMBI Goes to School: Evaluating Italian BabyLMs with Invalsi-ITA. In Proceedings of the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025) (pp. 1-12). Università di Cagliari.

Capone, L., Suozzi, A., Lebani, G. E., & Lenci, A. (2024). Babies: A benchmark for the linguistic evaluation of italian baby language models. In Proceedings of the 10th Italian Conference on Computational Linguistics (CLiC-it 2024) (pp. 157-170).

---