List your company

Commonsense Reasoning on HellaSwag (accuracy, retention)

59.9AccuracyTeacher FP16
23.60433.02742.4551.873Apr 28, 2026

Evaluation results

MethodLinks
Teacher FP16Size=1.3B, Precision=FP16, Evaluation Protocol=Zero-shot2026.0459.9—
BitMamba-2Size=1B, Precision=1.58-bit, Evaluation Protocol=Zero-shot2026.0445.6—
DF-SSMSize=278 MB, Evaluation Protocol=Zero-shot2026.0441.469.1
RandomEvaluation Protocol=Zero-shot2026.0425—