bond005/ru_llm_calibration

该数据集由Ivan Bondarenko创建，用于校准（重要性矩阵计算）和评估针对俄语的大语言模型的GGUF量化，包括但不限于Meno-Lite-0.1-GGUF模型。数据集分为训练集和测试集：训练集包含俄语小说、官方/商务风格文本、精选的维基百科文章以及新西伯利亚国立大学网站的文本样本，用于llama.cpp量化的校准；测试集包含关于俄罗斯各大学的随机文本，用于通过llama-perplexity等工具进行质量评估（如困惑度）。数据集的限制包括领域偏差（过度代表学术、官方和维基百科风格的俄语，可能不反映非正式、方言或高度特定领域的语言）、无多轮对话或对话内容（仅适用于因果语言模型评估，不适用于对话或指令遵循基准）以及时间截止（基于静态快照，可能不反映最新的语言使用或事件）。

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. It consists of a train split for calibration with llama.cpp quantization, containing Russian texts of fiction and official/business styles, selected Wikipedia articles, and sampled textual content from the Novosibirsk State University website; and a test split for quality evaluation (e.g., perplexity) via llama-perplexity or similar tools, containing randomly selected texts about various Russian universities. Limitations include domain bias (overrepresenting academic, official, and Wikipedia-style Russian, potentially lacking informal, dialectal, or highly domain-specific language), no multi-turn or dialogue (suitable only for causal LM evaluation, not conversational or instruction-following benchmarks), and temporal cutoff (based on static snapshots, may not reflect recent language usage or events).