sbert-training-config-snli-multinli
IN premise — summaries/2026/08/24/reimers-2019-sentence-bert-s3-model.md
Created 2026-08-25T02:58:28+00:00
SBERT is trained on SNLI (570K pairs) + MultiNLI (430K pairs) with 3-way softmax classification, 1 epoch, batch size 16, Adam optimizer at lr=2e-5, linear warm-up over 10% of steps, MEAN pooling.
Summary
This records the exact recipe used to train the sentence-embedding model: it learns to judge whether two sentences entail, contradict, or are neutral to each other by going through about one million labeled pairs once, using a specific optimizer, learning rate, and pooling strategy. Every downstream task that depends on how well the model captures meaning or similarity between sentences ultimately rests on this particular training setup being the right one.