Kurdish Interpretability
Mechanistic interpretability for low-resource Kurdish language models - KI mal anders
- Date
- since 2024
- Status
- active
Summary
Applying mechanistic interpretability methods to language models trained on Kurdish - AI differently, KI mal anders ;)
Overview
Kurdish Interpretability - abbreviated KI, with a deliberate nod to the German word for AI - applies mechanistic interpretability methods to language models trained on or applied to Kurdish. The central question is: what do these models actually learn about Kurdish, and more specifically, about its dialectal variation? Do models encode dialect as a coherent feature, or do they represent it as noise?
Why Interpretability for Low-Resource Languages?
Most mechanistic interpretability research has been conducted on large English-language models. Low-resource languages present a qualitatively different regime: models trained on them have seen far less data, may rely more heavily on transfer from other languages, and may have learned different internal representations as a result. Understanding these representations is both scientifically interesting and practically important - it can reveal failure modes invisible from benchmark scores alone.
Research Questions
- Which components of a multilingual model encode Kurdish-specific linguistic features?
- Is dialectal variation represented as a continuous latent dimension, a discrete cluster, or something else?
- How does the model's internal representation of Kurdish change as more Kurdish training data is provided?
- Can interpretability findings guide more targeted fine-tuning strategies for low-resource languages?
- [TODO: refine research questions as the project develops]
Current Status
This project is in the exploratory phase. Initial experiments are underway. Results and publications will be added here as the work progresses. [TODO: update with experiments and findings]
Keywords
- mechanistic interpretability
- low-resource
- Kurdish
- language models
- explainability