…oading

* Prototype ppo + ray

* reduce gradient

* push

* push changes

* quick push

* cache changes; this actually works with 6 nodes

* push changes

* psuh changes

* push the latest change

* push changes

* Fix uploading

* Make style

* style and quality

* update docs

* update mason.py

* log wandb tables

* update docs

* make style quality

* make sure to save the right thing

* push changes

* push

* push

* push changes

* push

* push

* fix

* remove preemption code

* fix

* push changes

* push

* quick fix

* quick push

* add ability to use alternate image

* rever

* final configs

* update

* unseen evals

* default value

* conditional

* flag for unseen evals

* safety evals and parquet files

* import Dataset

* update data dist plots

* nit

* smooth operation

* updates

* nits

* clean git

* cleaning for final SFT version

* oe eval priority

* up

---------

Co-authored-by: Nathan Lambert <[email protected]>

* fix and add script

* update

* fix copilot typo

* Reorganize the data preparation scripts for tulu v1 and v2.

* Minor improvement

* Remove open_platypus_commercial subset from Daring-Anteater

* Use hard-coded examples repo.

* Fix some bugs.

* Add OpenMathInstruct.

* Add a few more v3.5.x SFT mix ablations for the cleaner datasets.

* More experiments on mixes.

* help merge

* prep for merge

* reapply changes

* fix naming

---------

Co-authored-by: Nathan Lambert <[email protected]>

* Use vllm for all evaluations

* Do not use VLLM only for MMLU and TruthfulQA

* Quick change

* weight converter

* Support weka eval

* quick fix

* Quick change

* weight converter

* Add olmo1124 converter