license: mit
task_categories:
- feature-extraction
tags:
- multilingual
- llm
- linguistics
- embeddings
This dataset contains the computed language latent vectors (binary vectors, Euclidean vectors, and distances) as presented in the paper Deep Language Geometry: Constructing a Metric Space from LLM Weights.
The paper introduces a novel framework that utilizes the internal weight activations of Large Language Models (LLMs) to construct a metric space of languages. This dataset makes the automatically derived high-dimensional vector representations for 106 languages publicly available, capturing intrinsic language characteristics that reflect linguistic phenomena.
Paper: Deep Language Geometry: Constructing a Metric Space from LLM Weights
Code: https://github.com/mshamrai/deep-language-geometry
Gradio Analysis Tool (Hugging Face Space): https://huggingface.co/spaces/mshamrai/language-metric-analysis
Dataset Contents
The dataset includes:
- Calculated binary vectors
- Euclidean vectors
- Distances between languages
These components can be used to analyze and visualize inter-language connections and linguistic families.