Hello OpenUnlearning maintainers,
Thank you for releasing the OpenUnlearning framework and the TOFU
checkpoints. We are conducting an academic audit of representation-level
residual knowledge after unlearning.
We are currently using the following two public models:
Full model:
open-unlearning/tofu_Llama-3.2-1B-Instruct_full
RMU model:
open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_RMU_lr2e-05_layer5_scoeff1_epoch10
We have completed a static compatibility audit. The two repositories contain
complete model weights rather than adapters, and their architecture,
tokenizer, configuration, state-dict keys, tensor shapes, and dtypes are
compatible.
Before using the RMU checkpoint in formal experiments, we would appreciate
clarification on the following points:
-
What license applies specifically to the released RMU model weights?
Does the checkpoint inherit the Llama 3.2 Community License or have any
additional usage or redistribution restrictions?
-
Was this RMU checkpoint initialized directly from
open-unlearning/tofu_Llama-3.2-1B-Instruct_full?
If so, which exact revision of the Full model was used?
-
Which OpenUnlearning commit, Hydra configuration, training command,
random seed, and dataset revisions produced this checkpoint?
-
Could you clarify the exact meanings of:
- lr2e-05
- layer5
- scoeff1
- epoch10
- the internal config value _name_or_path=checkpoint-130
-
Is the Hugging Face checkpoint the original training output, or was it
merged, converted, or re-saved before release?
-
Is there a model card, experiment log, or provenance manifest that we
should cite when reporting results obtained with this checkpoint?
Our immediate goal is only an internal six-sample representation-interface
smoke test. We will avoid claiming a controlled before/after causal comparison
until the checkpoint provenance is confirmed.
Thank you for your time and for making these resources available.
Hello OpenUnlearning maintainers,
Thank you for releasing the OpenUnlearning framework and the TOFU
checkpoints. We are conducting an academic audit of representation-level
residual knowledge after unlearning.
We are currently using the following two public models:
Full model:
open-unlearning/tofu_Llama-3.2-1B-Instruct_full
RMU model:
open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_RMU_lr2e-05_layer5_scoeff1_epoch10
We have completed a static compatibility audit. The two repositories contain
complete model weights rather than adapters, and their architecture,
tokenizer, configuration, state-dict keys, tensor shapes, and dtypes are
compatible.
Before using the RMU checkpoint in formal experiments, we would appreciate
clarification on the following points:
What license applies specifically to the released RMU model weights?
Does the checkpoint inherit the Llama 3.2 Community License or have any
additional usage or redistribution restrictions?
Was this RMU checkpoint initialized directly from
open-unlearning/tofu_Llama-3.2-1B-Instruct_full?
If so, which exact revision of the Full model was used?
Which OpenUnlearning commit, Hydra configuration, training command,
random seed, and dataset revisions produced this checkpoint?
Could you clarify the exact meanings of:
Is the Hugging Face checkpoint the original training output, or was it
merged, converted, or re-saved before release?
Is there a model card, experiment log, or provenance manifest that we
should cite when reporting results obtained with this checkpoint?
Our immediate goal is only an internal six-sample representation-interface
smoke test. We will avoid claiming a controlled before/after causal comparison
until the checkpoint provenance is confirmed.
Thank you for your time and for making these resources available.