Language Modeling on Enron Dataset
3.4PerplexityPruning
Evaluation results
| Method | Links | ||
|---|---|---|---|
| PruningBackbone=Llama-2-7b2026.02 | 3.4 | 1.22 | |
| OriginalBackbone=Llama-2-7b2026.02 | 3.42 | 1.23 | |
| GhostBackbone=Llama-2-7b2026.02 | 5.64 | 1.73 | |
| PruningBackbone=Llama-3-8b2026.02 | 6.39 | 1.85 | |
| PruningBackbone=Llama-3.1-8b2026.02 | 6.39 | 1.89 | |
| OriginalBackbone=Llama-3-8b2026.02 | 6.45 | 1.86 | |
| OriginalBackbone=Llama-3.1-8b2026.02 | 6.53 | 1.88 | |
| PruningBackbone=GPT-22026.02 | 10.82 | 2.4 | |
| GhostBackbone=Llama-3.1-8b2026.02 | 10.89 | 2.39 | |
| GhostBackbone=Llama-3-8b2026.02 | 11.03 | 2.4 | |
| OriginalBackbone=GPT-22026.02 | 11.05 | 2.4 | |
| FFTParams=124.0M, Privacy Budget (ε)=Non-Priv.2026.01 | 14.31 | — | |
| LoRAParams=497.7K, Privacy Budget (ε)=Non-Priv., Aggregation=average over ranks2026.01 | 18.79 | — | |
| GhostBackbone=GPT-22026.02 | 19.12 | 2.95 | |
| TTLoRAParams=65.6K, Privacy Budget (ε)=Non-Priv., Aggregation=average over ranks2026.01 | 20.14 | — | |
| LoRA (avg.)Privacy Budget (epsilon)=3.0, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.71 | — | |
| LoRA (avg.)Privacy Budget (epsilon)=5.0, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.71 | — | |
| LoRA (avg.)Privacy Budget (epsilon)=0.5, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.72 | — | |
| LoRA (avg.)Privacy Budget (epsilon)=1.0, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.72 | — | |
| TTLoRA (avg.)Privacy Budget (epsilon)=0.5, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.72 | — | |
| TTLoRA (avg.)Privacy Budget (epsilon)=1.0, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.72 | — | |
| TTLoRA (avg.)Privacy Budget (epsilon)=3.0, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.72 | — | |
| TTLoRA (avg.)Privacy Budget (epsilon)=5.0, Training=DP-SGD, Ranks averaged=2-162026.01 | 20.72 | — | |
| FFTPrivacy Budget (epsilon)=5.0, Training=DP-SGD2026.01 | 20.9 | — | |
| FFTPrivacy Budget (epsilon)=3.0, Training=DP-SGD2026.01 | 20.99 | — | |
| FFTPrivacy Budget (epsilon)=1.0, Training=DP-SGD2026.01 | 21.25 | — | |
| FFTPrivacy Budget (epsilon)=0.5, Training=DP-SGD2026.01 | 21.48 | — | |
| FFTParams=124.0M, Privacy Budget (ε)=5.02026.01 | 24.92 | — | |
| TTLoRAParams=65.6K, Privacy Budget (ε)=5.0, Aggregation=average over ranks2026.01 | 24.99 | — | |
| TTLoRAParams=65.6K, Privacy Budget (ε)=3.0, Aggregation=average over ranks2026.01 | 25.37 | — | |
| FFTParams=124.0M, Privacy Budget (ε)=3.02026.01 | 25.38 | — | |
| LoRAParams=497.7K, Privacy Budget (ε)=5.0, Aggregation=average over ranks2026.01 | 26.44 | — | |
| TTLoRAParams=65.6K, Privacy Budget (ε)=1.0, Aggregation=average over ranks2026.01 | 26.55 | — | |
| LoRAParams=497.7K, Privacy Budget (ε)=3.0, Aggregation=average over ranks2026.01 | 26.64 | — | |
| FFTParams=124.0M, Privacy Budget (ε)=1.02026.01 | 27.05 | — | |
| TTLoRAParams=65.6K, Privacy Budget (ε)=0.5, Aggregation=average over ranks2026.01 | 27.52 | — | |
| LoRAParams=497.7K, Privacy Budget (ε)=1.0, Aggregation=average over ranks2026.01 | 27.62 | — | |
| FFTParams=124.0M, Privacy Budget (ε)=0.52026.01 | 28.81 | — | |
| LoRAParams=497.7K, Privacy Budget (ε)=0.5, Aggregation=average over ranks2026.01 | 29.29 | — |