Scale collaborates with NVIDIA with NeMo SteerLM

Scale AI is collaborating with NVIDIA to create the next generation of generative AI models using NVIDIA's NeMo SteerLM technique and Scale's high-quality training datasets.

What's going on here?

SteerLLM from Nvidia and Scale AI will help developers make apps more aligned towards their outcomes.

What does this mean?

SteerLM comes with an open-source evaluation dataset with 37k samples and different dimensions like helpfulness, correctness, complexity and verbosity.

SteerLM allows developers to dynamically customize model behaviour through easy-to-adjust attributes instead of full retraining. For example, in education applications, SteerLM can tailor model complexity and verbosity to individual learning needs. In gaming, SteerLM enables shaping NPC personality and emotional range.

Why should I care?

Making LLMs perform well in industry-specific scenarios is the target for many providers at the moment. Two of the major roadblocks in achieving that are a) access to industry-specific, custom-labelled, high-quality data and b) readymade frameworks with options to crank up/down certain behaviours.

Scale and Nvidia’s partnership is one solution for both of these. I’d bet on getting more of such options to come out of LLM fine-tuning/deployment companies in the next months.