List your company

Instruction Following on Instruction Following SFT 1.0 (eval)

59.4SFT ScoreRandom Sampling KD (Ours 12)
54.30455.62756.9558.273Mar 21, 2025

Evaluation results

MethodLinks
Random Sampling KD (Ours 12)Unique Tokens=122025.0359.4
Random Sampling KD (Ours 12+)Unique Tokens=12, Adaptive Weighting=true2025.0359.1
FullKD2025.0358.4
Top-K 50K=502025.0358.3
Top-K 12K=122025.0357.7
CE2025.0354.5