Skip to content

Dataset · eaddario

Imatrix calibration dataset for LLaMA.cpp

eaddario/imatrix-calibration

Small text dataset for calculating the importance matrix during model quantization.

What's inside

This is calibration data for generating the importance matrix (imatrix), which is needed to reduce errors when quantizing models in LLaMA.cpp. The llama-imatrix program does not read parquet directly, so it is suggested to convert the files to a text format, for example via DuckDB. The card specifies the languages ar, de, en, es, fr, hi and a size of less than 1k examples.

What you can do with it

  • Imatrix calibration before LLM quantization in LLaMA.cpp
  • Preparing a text file for llama-imatrix
  • Testing the quality of quantized models on multilingual texts
Tasks
Text generation
Languages
Arabic, German, English, Spanish, French, Hindi, Indonesian, Italian and 4 more
Size
up to 1K rows
Downloads this month
9K
Likes
79

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("eaddario/imatrix-calibration")