List your company
ResearchFormerly WizWand

Natural Language Inference on ANLI R2 (Accuracy)

81.14AccuracyIn-Squeeze
17.450433.985250.5267.0548Mar 21, 2022Nov 13, 2022Jul 8, 2023Mar 1, 2024Oct 24, 2024Jun 18, 2025Feb 11, 2026
Updated 5mo ago

Evaluation results

MethodLinks
In-Squeezerank=128 to 1, schedule=Min steps2026.0281.14
In-Squeezerank=128 to 1, schedule=Standard2026.0279.63
Cont-Squeezerank=128 to 1, additional_steps=7002026.0279.52
Direct Fine-tuningrank=1, additional_steps=02026.0279.3
Direct Fine-tuningrank=1, additional_steps=7002026.0278.57
Cont-Squeezerank=128 to 1, additional_steps=2002026.0278.46
Direct Fine-tuningrank=1, additional_steps=2002026.0277.23
Direct Fine-tuningrank=1, training_steps=+0 steps2026.0275.11
Direct Fine-tuningrank=1, training_steps=+200 steps2026.0274.72
In-Squeezereduction=128 ... -> 1, strategy=Min steps2026.0274.33
In-Squeezereduction=128 ... -> 1, strategy=Standard2026.0274.16
Cont-Squeezereduction=128 -> 1, training_steps=+200 steps2026.0273.66
Cont-Squeezereduction=128 -> 1, training_steps=+700 steps2026.0272.88
Direct Fine-tuningrank=1, training_steps=+700 steps2026.0272.38
Self-consistencyPrompting strategy=Self-consistency, Backbone model=PaLM-540B2022.0364.5
PaLM 2-Lprompting=1-shot2023.0563.4
CoT-promptingPrompting strategy=Chain-of-Thought, Backbone model=PaLM-540B2022.0358.9
Standard-prompting (no-rationale)Prompting strategy=no-rationale, Backbone model=PaLM-540B2022.0355.8
PaLM 2-Mprompting=1-shot2023.0549.5
PaLM 2-Sprompting=1-shot2023.0548.8
PaLMprompting=1-shot2023.0548.7
DenseBase Model=LLaMA-2-7B, Sparsity=Dense2025.0637.2
DenseBase Model=DeepSeek-7B, Sparsity=Dense2025.0636.6
SparsegptBase Model=LLaMA-2-7B, Sparsity=2:42025.0634.1
MaskProBase Model=LLaMA-2-7B, Sparsity=2:42025.0634.1
Pruner-ZBase Model=LLaMA-2-7B, Sparsity=2:42025.0633.9
SparsegptBase Model=DeepSeek-7B, Sparsity=2:42025.0633.7
MaskProBase Model=DeepSeek-7B, Sparsity=2:42025.0633.5
Pruner-ZBase Model=DeepSeek-7B, Sparsity=2:42025.0633.2
Cont-Squeezerank=128 to 1, additional_steps=02026.0229.3
Cont-Squeezereduction=128 -> 1, training_steps=+0 steps2026.0221.26
Zero-shotmode=0-S2026.0219.9