sergeyzh/rubert-mini-sts
Full fine-tune (fewer layers) vs cointegrated/rubert-tiny2. cointegrated/rubert-tiny2 matches part of sergeyzh/rubert-mini-sts (62 of its 126 tensors; the rest of sergeyzh/rubert-mini-sts has no counterpart). On the shared part: Fine-tune: 97% of the weights changed, across 3 of 3 layers.
Tensors changed: Embeddings 4/5, Norms 12/14, Attention 24/24, other 14/19
2 derivatives checked
- sergeyzh/rubert-mini-frida: Full fine-tune
- sergeyzh/rubert-mini-uncased: Different architecture