Views
No views yet
muse-bench/MUSE-News_target unlearned with method npo_gdr (10 epochs, alpha=1, lr 1e-5, max_len 2048) using the FailureLLMUnlearning (ICLR'25 Catastrophic Failure of LLM Unlearning via Quantization) base code. Tokenizer: meta-llama/Llama-2-7b-hf.| precision | verbmem_f | privleak | knowmem_f | knowmem_r |
|---|---|---|---|---|
| full | 0.00 | 104.63 | 45.46 | 34.71 |
| nf4 | 11.93 | 76.72 | 47.52 | 47.49 |
| gptq4 | 6.20 | 82.75 | 43.47 | 42.14 |
| awq4 | 18.31 | -93.82 | 45.73 | 40.42 |
full = bf16 (this repo); nf4 = bitsandbytes NF4 load-time; gptq4 = Jeesup/MUSE-News_Llama-2-7b_npo_gdr_alpha1_ep10_gptq4; awq4 = Jeesup/MUSE-News_Llama-2-7b_npo_gdr_alpha1_ep10_awq4.Jeesup/MUSE-News_Llama-2-7b_unlearning_results.