How to interpret the TOFU result

[zijianh4]

Hi,
Thanks for your great work. While I understand this github is still in development, could you please tell me how to interpret the TOFU result in https://github.com/locuslab/open-unlearning/blob/main/docs/results.md? I'm confused since I saw the forget_quality is 0 for the model finetuned with Retain data, which means that this uses the log p-value, but this conflicts with the positive values in other places. Could you please help to clarify this? Thanks!

[Dornavineeth]

Hi! Thank you for pointing this out. This is an error while drafting the readme. Updated it now!

NOTE: When you set retain_logs_path during evaluation, forget_quality is calculated or else it will be set to None.

Here is the command to set it:

python src/eval.py --config-name=eval.yaml experiment=eval/tofu/default \
  model=Llama-3.2-1B-Instruct \
  model.model_args.pretrained_model_name_or_path=open-unlearning/tofu_Llama-3.2-1B-Instruct_full \
  task_name=SAMPLE_EVAL \
  retain_logs_path=saves/eval/tofu_Llama-3.2-1B-Instruct_retain90/TOFU_EVAL.json

Please follow the Data Setup steps to download the metrics for all the retain models we provide.

[molereddy]

If the updates have resolved the issue, please close it or let us know of any further issues/clarifications.

Thank you!

[zijianh4]

Hi,
Sorry to interrupt again. Is there probably a bug in the current implementation in DPO? When I try running the DPO unlearning in the tofu_unlearn.sh file, there is an error KeyError: 'original', where I think the the inputs["forget"] only contains ['input_ids', 'attention_mask', 'labels']. Could you please help me to resolve this? Thanks!

[Dornavineeth]

Hi, thanks for raising this issue! It’s now fixed.

DPO methods require a different dataset that returns items with keys original and alternate, which is utilized in DPO trainer. You can check the QAwithIdkDataset class, which implements this functionality.

In our experiments, we specify this dataset in the idk.yaml config file, which differs from the default setting in default.yaml.

Please re-run python setup_data.py for downloading the JSONL file with different "I Don't Know" responses.