openpecha/tibetan-metadata-llm-sft
收藏资源简介:
这是一个藏文元数据LLM监督微调数据集(10%试点版本),是完整TiLamb SFT JSONL的10%分层随机子样本,用于试点LoRA训练和冒烟测试。该数据集用于从BDRC大纲段中提取标题和作者范围的监督微调,专为TiLamb-7B模型设计。数据集包含title和author的训练、验证和测试文件,每个训练行包括固定的任务提示(标题或作者)、裁剪后的段文本(通过TiLamb分词器限制在≤3584个标记内)以及输出JSON格式的跨度信息(包含文本、起始和结束位置,相对于裁剪后的偏移)。子采样方法为每个分割文件随机抽取10%的行(种子42)。裁剪策略分为三种:完整(整个段适合标记预算)、正例(包含黄金跨度的随机窗口,以抗位置偏差)和负例(随机窗口,该任务为空跨度)。数据集基于ganga4364/tibetan-metadata-extracted构建,使用OpenPecha/tibetan-text-meta-detection的llm_sft包开发。
This is a Tibetan metadata LLM supervised fine-tuning (SFT) dataset (10% pilot version), which is a 10% stratified random subsample of the complete TiLamb SFT JSONL, intended for pilot LoRA training and smoke testing. This dataset is designed for supervised fine-tuning to extract titles and author ranges from BDRC outline segments, and is specifically developed for the TiLamb-7B model. The dataset includes training, validation, and test files for title and author extraction tasks. Each training line contains a fixed task prompt (for title or author), cropped segment text limited to ≤3584 tokens via the TiLamb tokenizer, and output JSON-formatted span information, which includes the text, start and end positions relative to the cropped offset. The subsampling method is to randomly select 10% of lines from each split file with a random seed of 42. The cropping strategies are divided into three types: full (the entire segment fits within the token budget), positive example (a random window containing the gold span to mitigate position bias), and negative example (a random window where the task has no valid span). This dataset is built upon ganga4364/tibetan-metadata-extracted, and developed using the llm_sft package from OpenPecha/tibetan-text-meta-detection.




