Towards Cross-Tokenizer Distillation: the Universal Logit Distillation Loss for LLMs
Paper
•
2402.12030
•
Published
LLMs Distillation
The ULD loss, based on optimal transport, enables distillation across different LLM families without requiring shared tokenizers.
-
-
mistralai/Mistral-7B-Instruct-v0.2
Text Generation • Updated • 1.61M • • 2.75k -
meta-llama/Llama-2-7b-chat-hf
Text Generation • Updated • 1.23M • 4.4k -
EleutherAI/pythia-160m-deduped
Text Generation • Updated • 44.9k • 3 -
EleutherAI/pythia-410m-deduped
Text Generation • Updated • 16.6k • 20 -
EleutherAI/pythia-1b-deduped
Text Generation • Updated • 24.4k • 19 -
bigscience/bloomz-560m
Text Generation • Updated • 880k • 122 -
bigscience/mt0-base
Text2Text Generation • Updated • 3.19k • 31 -
facebook/opt-350m
Text Generation • Updated • 177k • 143 -
rajpurkar/squad
Viewer • Updated • 98.2k • 63.5k • 303 -
google-research-datasets/qed
Updated • 65 • 3 -
GEM/FairytaleQA
Viewer • Updated • 10.6k • 155 • 8 -
qiaojin/PubMedQA
Viewer • Updated • 274k • 12.2k • 218 -
knkarthick/dialogsum
Viewer • Updated • 14.5k • 30.6k • 197