Model Editing on Memorized Sequences subsampled for comparability (train)
9Skip RateEOS
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| EOSModel=Llama2-7B, Params=6.7B, k=5, beta=1002026.06 | 9 | 69.6 | 0.15 | |
| Next-bestModel=Llama2-7B, Params=6.7B, k=5, beta=1002026.06 | 9 | 68.3 | 0.15 | |
| SuppressModel=Llama2-7B, Params=6.7B, k=5, beta=1002026.06 | 9 | 49 | 0.15 | |
| RedactModel=Llama2-7B, Params=6.7B, k=5, beta=1002026.06 | 9 | 40.7 | 0.15 | |
| EOSModel=OLMo-7B, Params=6.9B, k=5, beta=1002026.06 | 10.3 | 85.7 | 0.41 | |
| Next-bestModel=OLMo-7B, Params=6.9B, k=5, beta=1002026.06 | 10.3 | 81.5 | 0.41 | |
| SuppressModel=OLMo-7B, Params=6.9B, k=5, beta=1002026.06 | 10.3 | 60.5 | 0.41 | |
| RedactModel=OLMo-7B, Params=6.9B, k=5, beta=1002026.06 | 10.3 | 31.7 | 0.41 | |
| EOSModel=OLMo-1B, Params=1.2B, k=5, beta=1002026.06 | 43.4 | 52.4 | 0.04 | |
| Next-bestModel=OLMo-1B, Params=1.2B, k=5, beta=1002026.06 | 43.4 | 55.4 | 0.04 | |
| SuppressModel=OLMo-1B, Params=1.2B, k=5, beta=1002026.06 | 43.4 | 36.2 | 0.04 | |
| RedactModel=OLMo-1B, Params=1.2B, k=5, beta=1002026.06 | 43.4 | 15.7 | 0.04 | |
| EOSModel=SmolLM-360M, Params=360M, k=5, beta=1002026.06 | 48.3 | 34.9 | 0.04 | |
| Next-bestModel=SmolLM-360M, Params=360M, k=5, beta=1002026.06 | 48.3 | 37.7 | 0.04 | |
| SuppressModel=SmolLM-360M, Params=360M, k=5, beta=1002026.06 | 48.3 | 30.6 | 0.04 | |
| RedactModel=SmolLM-360M, Params=360M, k=5, beta=1002026.06 | 48.3 | 15.3 | 0.04 |