Each dot is a prompt, placed by a 2D UMAP projection of its Harrier-OSS-v1-0.6B embedding, so nearby prompts are similar. Task types come from clustering these embeddings with UMAP and HDBSCAN, followed by manual screening.
Find prompts from SE research that are similar to yours
Paste a prompt you use, or describe the task you give the model. We compare it with the prompts mined from software engineering papers.
Semantic results are ranked by the cosine similarity of Harrier-OSS-v1-0.6B embeddings, the model used to cluster the dataset. Keyword results are ranked by TF-IDF.
- 1Describe your prompt
Paste a prompt, or a sentence on what you ask the model to do.
- 2We embed it
With Harrier-OSS-v1-0.6B, the model used to cluster the dataset, after stripping LaTeX and masking template slots.
- 3Compare
The closest of the prompts, with their paper, task type and place on the map.
Keyword mode matches words with TF-IDF instead. Prompts from papers under the default arXiv license are searchable, but their text is not redistributed, so those results link to the paper.
The people and plans behind SEPIA
SEPIA (Software Engineering Prompts In Articles) collects the prompts that software engineering researchers report in their papers, so that they can be found, compared and reused.
- Prompts
- –
- Papers
- –
- Task types
- –
- Years
- –
Team
-
Ali Al-KaswanDelft University of Technology
-
Razvan Mihai PopescuDelft University of Technology
-
Andy ZaidmanDelft University of Technology
-
Maliheh IzadiDelft University of Technology
Where to find SEPIA
-
Dataset on ZenodoAvailable
The prompts, paper records, task types and all human annotations, with the mining pipeline.
Open On ZenodoDOI 10.5281/zenodo.24094112 -
Hugging FaceAvailable
The dataset as a ready-to-load Hugging Face dataset.
Open On Hugging Facehuggingface.co/datasets/aalkaswan/sepia -
Code on GitHubComing soon
The mining pipeline, the clustering and this explorer, as an open repository.
-
PaperComing soon
How we built the dataset, how well the detector works, and what the prompts show about SE research.
What we plan next
- PDF-only papers
Extract prompts from the cs.SE papers that are on arXiv only as PDF.
- SE venues beyond arXiv
Cover papers from software engineering venues that are not posted on arXiv.
- Images and replication packages
Capture the prompts that papers show only as figures or keep in their artifacts.
- Prompting techniques
Annotate the techniques the prompts use, such as few-shot examples or chain-of-thought instructions.
- Better clustering
Improve how prompts are grouped into task types, so that fewer of them stay unassigned.
- More typed prompts
Validate the suggested task types for unassigned prompts and fold the reliable ones into the labels.
Contact
Questions about the dataset, a prompt we missed, a task type that looks wrong, or an idea for a collaboration? Write to us.
- Ali Al-Kaswana.al-kaswan@tudelft.nl
- Razvan Mihai Popescur.m.popescu@tudelft.nl
Delft University of Technology · Delft, The Netherlands
