Integrating gender inclusivity into large language models
summary
Addresses gender bias in Polish large language models and presents a concrete corrective approach: creating the IPIS dataset (~23,000 proofreading examples) and using LoRA instruction tuning on multilingual and Polish models (Llama-8B, Mistral-7B, Mistral‑Nemo, Bielik, PLLuM). Reports baseline small models ~0.20 F1 for inclusive forms, Bielik ~0.40, and after tuning one Bielik variant reaches ≈0.60 F1; notes translation data is smaller and can hurt performance, and the data plus a public competition are released for reproduction.
AI-written summary · about
chapters (8)
- 0:00Gender inclusivity in LLMs
- 1:07Polish language basics
- 9:59Masculinities and consequences
- 13:26Bias reduction methods
- 16:22Inclusive language models research
- 25:44Experiment results and challenges
- 26:28Polish inclusive LLM contest
- 27:32Summary and Q&A
related talks
16mBest Practices for Open Multilingual LLM Evaluation
39mBeyond Bias: Crafting Fairness in Pythonic Machine Learning Systems
31mCombining behavioral research and NLP to combat gender biases
30mBiases in Language Models
42mPost-Training Small Language Models: the adventures of a practitioner
01h 38mTraining Large Language Models: Practices and Research Questions
28mFine-Tuning Small Language Models with Python
31mCombining behavioral research & NLP to combat gender biases