marin-community/hero-run-4-code-sdg-prompts-python-fenced-n16
收藏资源简介:
该数据集是一个专门为Marin合成数据生成任务设计的提示源数据集,名为Hero Run 4 Code SDG Prompts (16x)。它包含Python代码提示,每个提示重复16次,总共有146,688行数据。数据集来源于mlfoundations-dev/hero_run_4_code,包含9,168个独特的提示。每个提示的生成顺序与源文件中首次出现的顺序一致,并且每个提示的响应索引从0到15连续排列。数据集包含多个列,如prompt_index、response_index、_unique_row_id、instruction_seed和generation_prompt。提示后缀要求用户将完整解决方案放在一个Python代码块中,并且不要在代码块后写任何文字。预期用途是使用generation_prompt作为模型输入列,并将模型生成的文本写入generated_text列。
This dataset is a prompt-only source dataset for Marin synthetic data generation jobs, named Hero Run 4 Code SDG Prompts (16x). It contains Python fenced-code prompts, with each prompt repeated 16 times, totaling 146,688 rows. The dataset is sourced from mlfoundations-dev/hero_run_4_code and includes 9,168 unique prompts. The order of prompts follows the first-seen order of unique instruction_seed values in the source parquet files, with response indices running contiguously from 0 to 15 for each prompt. The dataset includes columns such as prompt_index, response_index, _unique_row_id, instruction_seed, and generation_prompt. The prompt suffix instructs users to place the complete solution in a single Python fenced code block and to avoid writing any prose after the code block. The intended use is to employ generation_prompt as the model input column and write model completions to generated_text.




