The proposed logits-based fine-tuning, which mixes teacher logits with ground truth labels, improves math reasoning accuracy of small LLaMA models over standard supervised fine-tuning, with a controlled GSM8K gain of 2.0 points and larger claimed gains on other benchmarks.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Logits-Based Finetuning
The proposed logits-based fine-tuning, which mixes teacher logits with ground truth labels, improves math reasoning accuracy of small LLaMA models over standard supervised fine-tuning, with a controlled GSM8K gain of 2.0 points and larger claimed gains on other benchmarks.