AutoSDT: Scaling Data-Driven Discovery Tasks Toward Open Co-Scientists

Name: AutoSDT: Scaling Data-Driven Discovery Tasks Toward Open Co-Scientists
Uploaded: 2026-10-04T02:56:00.493588+00:00
Duration: 439 s
Description: AutoSDT is an automatic pipeline that collects high-quality coding tasks from real-world scientific workflows to build a massive dataset called AutoSDT5K. This process uses LLMs to search for relevant code, filter it, and format it into instruction-code pairs across four scientific fields. Fine-tuning models on this data significantly boosts their performance in complex discovery benchmarks.

AutoSDT is an automatic pipeline that collects high-quality coding tasks from real-world scientific workflows to build a massive dataset called AutoSDT5K. This process uses LLMs to search for relevant code, filter it, and format it into instruction-code pairs across four scientific fields. Fine-tuning models on this data significantly boosts their performance in complex discovery benchmarks.

Open the full episode