List your company
ResearchFormerly WizWand

Text Reconstruction from Gradients on SST-2

100ROUGE-1DAGER
11.18434.24257.380.358Feb 17, 2022Jul 4, 2022Nov 19, 2022Apr 6, 2023Aug 22, 2023Jan 7, 2024May 24, 2024
Updated 7mo ago

Evaluation results

MethodLinks
DAGERModel=GPT-2_BASE, Batch size (B)=162024.0510094.6——
DAGERModel=LLaMa-2 (7B), Batch size (B)=162024.05100100——
DAGERModel=GPT-2_BASE, Batch size (B)=322024.0510093.4——
DAGERModel=LLaMa-2 (7B), Batch size (B)=322024.0510099.9——
DAGERModel=GPT-2_BASE, Batch size (B)=642024.0510085——
DAGERModel=LLaMa-2 (7B), Batch size (B)=642024.0599.999.9——
DAGERModel=LLaMa-2 (7B), Batch size (B)=1282024.0599.397.8——
LAMPcosArchitecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.0291.658.279.7—
LAMPL2+L1Architecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.0289.753.275.4—
LAMPL2+L1Architecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.0289.355.575.9—
LAMPcosArchitecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.0288.856.977.7—
LAMPL2+L1Architecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.0288.657.475.7—
LAMPcosArchitecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.0288.555.976.5—
LAMPcosArchitecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.0287.654.176.1—
TAGArchitecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.0280.816.859.1—
DLGArchitecture=BERT-LARGE, Batch size=1, Fine-tuned=false2022.0278.418.159—
TAGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.0275.618.957.4—
DLGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=false2022.0265.417.754.2—
TAGArchitecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.0243.59.440.9—
DLGArchitecture=TinyBERT6, Batch size=1, Fine-tuned=false2022.02425.439.6—
LAMPL2+L1Architecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.0241.610.939.3—
TAGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.02405.736.6—
DLGArchitecture=BERT-BASE, Batch size=1, Fine-tuned=true2022.02362.733.9—
DAGERModel=GPT-2_BASE, Batch size (B)=1282024.0514.64.3——
DAGERModel=GPT-2, Hardware=RTX 5060, Batch Size (B)=12026.03———3
DAGERModel=GPT-2, Hardware=RTX 5060, Batch Size (B)=22026.03———4
DAGERModel=GPT-2, Hardware=RTX 5060, Batch Size (B)=42026.03———9
DAGERModel=GPT-2, Hardware=RTX 5060, Batch Size (B)=82026.03———18
DAGERModel=GPT-2, Hardware=RTX 5060, Batch Size (B)=162026.03———31
SOMPBeam Width (Bw)=4, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=12026.03———6
SOMPBeam Width (Bw)=4, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=92026.03———9
SOMPBeam Width (Bw)=4, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=42026.03———13
SOMPBeam Width (Bw)=4, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=82026.03———22
SOMPBeam Width (Bw)=4, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=162026.03———30
SOMPBeam Width (Bw)=8, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=12026.03———7
SOMPBeam Width (Bw)=8, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=22026.03———11
SOMPBeam Width (Bw)=8, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=42026.03———14
SOMPBeam Width (Bw)=8, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=82026.03———26
SOMPBeam Width (Bw)=8, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=162026.03———37
SOMPBeam Width (Bw)=16, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=12026.03———8
SOMPBeam Width (Bw)=16, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=22026.03———12
SOMPBeam Width (Bw)=16, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=42026.03———30
SOMPBeam Width (Bw)=16, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=82026.03———40
SOMPBeam Width (Bw)=16, Model=GPT-2, Hardware=RTX 5060, Batch Size (B)=162026.03———52