Ethahtz/creative-gemma-stories

该数据集包含来自LLM（大型语言模型）的创意生成内容，具体通过creative_tasks管道（使用generate_infinite_chats.py脚本）或任何兼容的generations.jsonl文件生成。数据集主要用于文本生成任务，特别是评估模型在创意写作中的多样性和响应能力。其中一个关键配置是google-gemma-2-9b-it_dsshort_story_prompts_bkvllm_seed42_topp0.9，它使用google/gemma-2-9b-it模型，基于100个未详细说明的短故事提示（来自creative_tasks/data/short_story_prompts.json）生成内容。这些提示形式为Write a short story about {SUBJECT}.，主题涵盖多个类别，旨在探究模型如何处理最小化指定的创意提示。生成参数包括随机种子42、top_p=0.9、temperature=1.0和最大新令牌数1024，共生成10000行数据。数据集强调可重复性，每个运行配置都在generation_config.json中有详细记录。

This dataset contains LLM creative generations from the creative_tasks pipeline (using the generate_infinite_chats.py script) or any compatible generations.jsonl file. It is designed for text-generation tasks, specifically to evaluate the diversity and responsiveness of models in creative writing. One notable configuration is google-gemma-2-9b-it_dsshort_story_prompts_bkvllm_seed42_topp0.9, which uses the google/gemma-2-9b-it model and is based on 100 underspecified short-story prompts from creative_tasks/data/short_story_prompts.json. Each prompt follows the form Write a short story about {SUBJECT}., with subjects drawn from various categories to probe how models handle minimally-specified creative prompts. Generation parameters include seed=42, top_p=0.9, temperature=1.0, and max_new_tokens=1024, resulting in 10,000 rows of data. The dataset emphasizes reproducibility, with full per-run parameters documented in generation_config.json.