ceselder/loracle-ia-warmstart-v5

loracle-ia-warmstart-v5数据集是为LoRAcle管道的Warmstart SFT阶段设计的。它由`ceselder/loracle-ia-warmstart`和`ceselder/loracle-ia-RL`两个数据集的联合构建而成，排除了20个组织的公平评估集。数据集采用随机75/25的分割方式，其中75%（662个LoRAs）用于本数据集，25%用于`loracle-ia-RL-v5`的配对RL阶段。数据集包含1909行数据，涉及662个独特的LoRAs，每行数据包括lora_id、source、qa_type等字段。数据集的类别包括quirk、harmful_roleplay、benign_roleplay等，qa_types包括adv_probe_no_trigger_state、adv_probe_swap_check等。

The loracle-ia-warmstart-v5 dataset is designed for the Warmstart SFT stage of the LoRAcle pipeline. It is built from the union of `ceselder/loracle-ia-warmstart` and `ceselder/loracle-ia-RL`, excluding the 20-org fair-eval set. The dataset uses a random 75/25 split, with 75% (662 LoRAs) allocated to this dataset and 25% reserved for the `loracle-ia-RL-v5` paired RL stage. The dataset contains 1909 rows with 662 unique LoRAs, each row including fields such as lora_id, source, and qa_type. Categories in the dataset include quirk, harmful_roleplay, benign_roleplay, etc., and qa_types include adv_probe_no_trigger_state, adv_probe_swap_check, among others.