Skip to content

FIX Dedupe prompts in the BeaverTails, PKU-SafeRLHF and Aegis loaders - #3017

Merged
Roman Lutz (romanlutz) merged 5 commits into
microsoft:mainfrom
u7k4rs6:fix/dataset-dedupe
Oct 9, 2026
Merged

Roman Lutz (romanlutz) merged 5 commits into
microsoft:mainfrom
u7k4rs6:fix/dataset-dedupe

Conversation

@u7k4rs6

Copy link
Copy Markdown
Contributor

Fixes #3016.

Each loader now keeps one seed per prompt and merges the harm categories from every row that has it. BeaverTails also ORs the per-category flags. Aegis keeps the first row's id and labels in metadata. Filters still apply per row, so only rows that pass add their categories.

On the real HF data: BeaverTails goes from 166,382 to 14,402 seeds, PKU-SafeRLHF from 73,907 to 38,641, Aegis from 19,093 to 13,246. All three stay in the huge size bucket; I updated the size comments.

Added a merge test per loader (all three fail on main). tests/unit/datasets: 4960 passed.

@romanlutz Roman Lutz (romanlutz) self-assigned this Oct 9, 2026
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

@romanlutz Roman Lutz (romanlutz) left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just making a small test adjustment but otherwise this looks great

Roman Lutz (romanlutz) and others added 2 commits October 9, 2026 06:02
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
@romanlutz
Roman Lutz (romanlutz) merged commit dfe3cac into microsoft:main Oct 9, 2026
55 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

BeaverTails, PKU-SafeRLHF and Aegis loaders return the same prompt many times

2 participants