Skip to content

Conversation

@molereddy
Copy link
Collaborator

@molereddy molereddy commented Apr 7, 2025

What does this PR do?

metrics

  • adds 6 of MIMIR's MIA attacks: minK, minK ++, ref, zlib, loss, GradNorm
  • adds ES and EM metrics
  • simplifies and standardize MIA evals structure/code
  • adds holdout dataset so TOFU allows for MIA based privleak evaluations: created by generating IID examples with the same GPT-4 endpoint and original prompt from TOFU's original creation
  • generalises KS-test, rel diff metrics

updated documentation

misc

  • adds seed in train and eval

Fixes # (issue)
#70

Before submitting

  • This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
  • Have you gone through the contributions guide?
  • Are your changes documented? Read documentation guidelines here.

molereddy and others added 17 commits March 1, 2025 09:13
* testing commit

* Fixes

* cleanup
Fix tofu_unlearn.sh for IdKDPO method.
* IdkDPO script fix in tofu_unlearn.sh (locuslab#65)

* Fix hyperlinks in README
* Download I don't know data in setup_data.py
* Fix tofu_unlearn.sh for IdkDPO

---------

Co-authored-by: Anmol Mekala <49127549+molereddy@users.noreply.github.com>

* overwrite=True

* RMU added

* Fix ref model device

* ruff fix

* RMU updated

* Update rmu.py

* Update README.md: add RMU

* Added references and renamed functions

---------

Co-authored-by: Anmol Mekala <49127549+molereddy@users.noreply.github.com>
…on (#8)

* docs: updates, small corrections, re-formats

* modified ruff commands

* modified ruff commands

* CI/CD minor updates

* added contributing + leaderboard

* fix minor spelling misatkes

* docs: bunch of minor updates

* docs fixes

---------

Co-authored-by: molereddy <m.anmolreddy@gmail.com>
* update docs

* setup: add license

* Update results docs

* re-factor MIA probs extraction code

* Ensure model and tokenizer are compatible wrt vocab size

* Model-tokenizer vocab check: treat as warning

* Document metric usage

* Merge main properly

* pre-commit enforcement to be set only in dev mode

* revert vocab size warning

* Revert "revert vocab size warning"

This reverts commit eded904.

* Revert "pre-commit enforcement to be set only in dev mode"

This reverts commit 2bb5e19.

* small fixes: keep pre-commit only in dev mode prereqs, remove vocab size check

* privleak changes: change relative_auc handler to privleak. move min_k logic to metric file from utils

* Simplified privleak version

* Access key option for datasets

* feat: MIA attack implementation from mimir, pipelined into open-unlearning format

* Tokenwise, example wise log probs computation

* configs: add MIA new version of privleak support

* feat: privleak in TOFU

* More hyperparam details: bsz

* load and register mia metrics, cleanup old

* MIA init bug fixes + convert numpy scalars to floats

* fix: load ref model for ref mia

* misc: comments and nits

* fix: ref attack load to same device, log things

* describe standard mia score signage

* fix: handle division by 0

* standardize auc convention (fix muse's)
make relative diff metric separate and clear

* generalise the ks-test usage beyond forget quality

* update docs: repo updates, clarify example commands, small fixes

* fix: wrong command in readme

* fit gradnorm to conventions

* updated list of metrics

* Add MIA attack configs

* load holdout dataset from local path (must be updated)

* docs: Update MIA count

* remove hanging comments

* fixes: docs, comments etc. addressing comments

* Replace scrollable with dropdown

* Ruff style changes

* Clarify metrics kwargs, small docs updates

* Delete todos.txt

* Alert rendering

* Mark alerts again

* More alert rendering

* more rendering changes

* Add star history

* Make heading smaller

* Add holdout dataset to metrics, datasets and experiment configs

* Add holdout evals to scripts

* Fix cfg spell mistake (#9)

* fix spelling of holdout

* fix typo

* Add links documentation

* Update misc documentation

* Small docs updates

* Clean up links

* Clarify new target models

* Add ES, EM metrics + Minor fixes (#10)

* Add seed in eval

* Add ES, EM metrics

* Add tqdm in MIA

* Fix get_data

* Ruff format fix (pre-commit and code)

* Update docs

* Update Readme

* Evaluation and leaderboard modifications (#11)

* Remove different splits in leaderboard

* update default metrics

* Fix: Only required metrics in summary

* Minor fixes in docs and gitignore

---------

Co-authored-by: Vineeth <48151992+Dornavineeth@users.noreply.github.com>
Co-authored-by: Dornavineeth <vineethdorna@gmail.com>
@molereddy molereddy linked an issue Apr 7, 2025 that may be closed by this pull request
5 tasks
@molereddy molereddy merged commit 2e16546 into locuslab:main Apr 7, 2025
1 check 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.

[Feature Request] More documentations

2 participants