Released Sentinel S1: open models that tell you when they don't know 🛡️

#1
by Sentovara - opened

LoRA adapters for Olmo 3 7B and Olmo 3.1 32B that start a reply with UNSURE: when the model is likely wrong, so you know when to look it up.

7B: plain answers right 75% vs 20% for the base model, and it flags 94% of its own mistakes.
32B: 68% vs 28%, and with search it gets 45% of flagged questions right vs 11% guessing.
Maths and instruction-following unchanged. 9/9 pre-registered checks passed on both.

Try it: https://huggingface.co/spaces/Sentovara/Sentinel-S1-chat
Models: https://huggingface.co/Sentovara/Sentinel-S1-7B · https://huggingface.co/Sentovara/Sentinel-S1-32B

Questions or want the recipe? Ask in the Community tab.

Sign up or log in to comment