Datasetpublic

bond005/ru_llm_calibration

Ru LLM calibration This dataset is created by Ivan Bondarenko for calibrating (importance matrix computation) and evaluating GGUF quantizations of large language models targeting Russian language, including but not limited to Meno-Lite-0.1-GGUF. Purpose Train split: calibration for llama.cpp quantization (any Russian-focused LLM). Test split: quality evaluation (perplexity, etc.) via llama-perplexity or similar tools. Dataset Composition Train:… See the full description on the dataset page: https://huggingface.co/datasets/bond005/ru_llm_calibration.

sourceHugging Facemitupdated 5mo agoView on Hugging Face

1likes6downloads

Dataset Card

Ru LLM calibration

This dataset is created by Ivan Bondarenko for calibrating (importance matrix computation) and evaluating GGUF quantizations of large language models targeting Russian language, including but not limited to Meno-Lite-0.1-GGUF.

Purpose

—Train split: calibration for llama.cpp quantization (any Russian-focused LLM).
—Test split: quality evaluation (perplexity, etc.) via llama-perplexity or similar tools.

Dataset Composition

—Train: Russian texts of fiction and official/business styles, including selected Wikipedia articles and sampled textual content from the Novosibirsk State University (NSU) website.
—Test: randomly selected texts about various Russian universities.

Limitations

—Domain bias: the dataset overrepresents academic, official, and Wikipedia-style Russian. It may not reflect informal, dialectal, or highly domain-specific language (e.g., medicine, law, technical manuals).

—No multi-turn or dialogue: suitable only for causal LM evaluation, not conversational or instruction-following benchmarks.

—Temporal cutoff: based on static snapshots (no streaming updates), so it may not reflect very recent language usage or events.