glob21/GrandMaster-PRO-MAX

GrandMaster-PRO-MAX是一个大型高质量俄语指令调优数据集，专门用于训练模型遵循多种语言（主要是俄语）的指令并主要用俄语回答。该数据集是第一个非通过翻译英文模型回答生成的俄语SFT数据集，其助手回答完全由GPT-4-Turbo-1106根据用户原始指令生成。数据来源多样，包括Saiga、Tagengo、lmsys/lmsys-chat-1m、lmsys/lmsys-arena-human-preference-55k、合成数据、abacusai/SystemChat-1.1、d0rj/OpenHermes-2.5-ru和BAAI/Infinity-Instruct等，仅使用指令部分，回答全部重新生成。数据集包含约15.5万唯一指令-回答对，生成成本约4.5千美元，具有内置Chain-Of-Thought（CoT）能力，通过复杂提示实现结构化回答。特征包括source、conversation（含content和role）、prompt_tokens、answer_tokens、cluster、prompt_lang和answer_lang，训练集151822例，测试集3291例。语言分布广泛，提示涉及95种语言（俄语88781例，英语64287例等），回答涉及60种语言（俄语128149例，英语25411例等）。数据集已更新，添加了新数据并进行了去重处理，还包含基于intfloat/multilingual-e5-large的聚类信息（15149个唯一簇）和令牌统计。

GrandMaster-PRO-MAX is a large, high-quality Russian instruction-tuning dataset designed to train models to follow diverse instructions in multiple languages (primarily Russian) and respond mainly in Russian. It is the first Russian SFT dataset generated without translating answers from English models, with assistant responses entirely generated by GPT-4-Turbo-1106 based on original user instructions. Data sources include Saiga, Tagengo, lmsys/lmsys-chat-1m, lmsys/lmsys-arena-human-preference-55k, synthetic data, abacusai/SystemChat-1.1, d0rj/OpenHermes-2.5-ru, and BAAI/Infinity-Instruct, using only prompts and regenerating all answers. The dataset contains approximately 155,000 unique instruction-answer pairs, with a generation cost of about $4.5k, and features built-in Chain-Of-Thought (CoT) capability through complex prompting for structured responses. Features include source, conversation (with content and role), prompt_tokens, answer_tokens, cluster, prompt_lang, and answer_lang, with 151,822 examples in the training set and 3,291 in the test set. Language distribution is broad: prompts cover 95 languages (Russian 88,781, English 64,287, etc.), and answers cover 60 languages (Russian 128,149, English 25,411, etc.). The dataset has been updated with new data and deduplication, and includes clustering information (15,149 unique clusters) based on intfloat/multilingual-e5-large and token statistics.