Prompt Separation

Prompt-Separation Voice-typed podcast prompt transcripts decomposed into structured fields: discrete prompts (asks), a list of context chunks, and free-form host_notes. The dataset supports training a small model that, given a single voice-typed message, recovers the structured fields an AI host would consume — separating "what is the user actually asking?" from "what is the surrounding context?" from "how should the response be shaped?". Source Prompts come from… See the full description on the dataset page: https://huggingface.co/datasets/danielrosehill/Prompt-Separation.

About This Project

Prompt-Separation is a dataset published on Hugging Face that contains voice-typed transcripts from the My Weird Prompts podcast production pipeline, each decomposed into structured fields: discrete prompts (asks), a list of context chunks, and free-form host notes. Each row corresponds to one episode's raw user message, keyed by the episode identifier from the source Postgres database. The dataset is released under a CC-BY-4.0 license in English, stored in Parquet format with a CSV mirror, and uses a wide flat schema with up to ten prompt columns and twenty context columns.

The labels are a mix of hand-annotated rows and silver labels generated by an LLM (deepseek-v4-flash) using human-annotated rows as few-shot exemplars. Only ten rows are marked as human-annotated, so the silver labels are intended for weak-supervision training and bootstrapping rather than evaluation. The dataset documentation explicitly notes that the prompt fields are abstractive rewrites rather than verbatim spans of the transcript. The labelling pipeline itself is maintained in a separate private repository.

The dataset may be useful to researchers and developers working on text classification or text generation for prompt engineering, particularly those building small models that recover the structured fields an AI host would need from a single voice-typed message. It may also interest those studying weak supervision or dataset bootstrapping from LLM-generated labels.

Project Details

Tags

task_categories:text-classificationtask_categories:text-generationlanguage:enlicense:cc-by-4.0size_categories:1k<n<10kformat:parquetmodality:tabularmodality:textlibrary:datasetslibrary:pandaslibrary:polarslibrary:mlcroissantregion:usprompt-engineeringvoice-typedweak-supervisionprompt-separation

Explore More Projects