Granite-3.3-2B-Instruct

Gold definitionUpdated Apr 2, 2026

Definition

Granite-3.3-2B-Instruct is a 2-billion parameter instruction-tuned Large Language Model (LLM) from the Granite family. It was fine-tuned and evaluated in research exploring Reinforcement Learning with Verifiable Rewards (RLVR) to enhance intellectual humility and reduce hallucinations.

At a glance

Executive summary

Granite-3.3-2B-Instruct is a 2-billion parameter AI language model used in research to make models more honest. By training it to say 'I don't know' when uncertain, researchers aim to reduce false information and improve its reliability, especially in factual questions.

TL;DR

A 2-billion parameter AI model trained to be more reliable by learning to admit when it doesn't know an answer, reducing made-up information.

Key points

Fine-tuned with Reinforcement Learning with Verifiable Rewards (RLVR) using a ternary reward structure to incentivize abstention.
Addresses LLM hallucination and unreliability in factual domains by promoting intellectual humility.
Used by researchers and ML engineers focused on model safety, truthfulness, and responsible AI deployment, especially in critical applications like medical or mathematical QA.
Unlike standard supervised fine-tuning that primarily rewards correctness, RLVR explicitly incorporates abstention as a rewarded behavior.
Part of a broader trend towards developing more trustworthy and aligned AI models through advanced reinforcement learning and human feedback techniques.

Use cases

Medical Question Answering: Providing reliable answers to medical queries, with the ability to abstain rather than hallucinate potentially harmful information.
Financial Advisory Systems: Offering financial advice where accuracy is paramount, and the model can indicate uncertainty instead of giving incorrect recommendations.
Legal Document Review: Assisting in legal research by identifying relevant information and flagging areas where it lacks definitive knowledge, preventing misinterpretations.
Educational Tutoring: Guiding students by providing correct answers or admitting when it cannot confidently answer a complex question, fostering accurate learning.

Also known as

Granite-2B-Instruct, Granite-3.3B-Instruct, Granite LLM