Medium Foundation Models · 1 min read
Solar 10.7B: depth upscaling to merge layers from two LLaMA-2 models
In one sentence Upstage presents Solar 10.7B, created by merging intermediate layers of two fine-tuned LLaMA-2 models (depth upscaling), winning the MBTI-OpenLLM leaderboard in November 2023.
Reading level
Solar 10.7B comes from an unusual idea: take two fine-tuned LLaMA-2 models and combine them by joining their internal layers, like building a taller structure by taking floors from two different buildings.
This technique, called depth upscaling, produces a 10.7-billion-parameter model that performs better than the two original models separately.
Upstage's model won the open-source OpenLLM leaderboard in November 2023, proving competitive with models of much larger size.
Companies
Upstage
Tools
Solar-10.7B
Tags
SolarUpstageDepth UpscalingLLaMAModel Merging
Sources