Cohere Enhances its Large Language Models with Scale.
Cohere is Building the Future of Language AI
Cohere Needed High-Quality Data For Their Instruction-Following Model.
The team at Cohere had an urgent need to acquire high-quality data to train the first iteration of their instruction-following model, Command. However, the Cohere team, “needed to build up an annotation force quickly, and Scale provided access to a large-scale, diverse, and skilled workforce to employ for this complex data creation job,” explained Alex Wang, Technical Lead at Cohere. The team at Cohere knew they needed a partner who could provide diverse and complex inputs (prompts) and outputs (responses) to ensure their models would perform well on challenging and unusual queries their users may have.
Scale's Generative AI Data Engine Delivers High-Quality Prompt/Response Pairs.
Large Language Models (LLMs) are typically developed in 3 stages:
- Stage 1: Training with a large corpus of raw data
- Stage 2: Instruction fine-tuning using prompt/response pairs
- Stage 3: Reinforcement learning with human feedback (RLHF)
Scale and Cohere initially partnered on stage 2: instruction fine-tuning using prompt/response pairs. Partnering with Scale solved both of Cohere’s challenges. “They have a skilled workforce capable of delivering high-quality prompt/response pairs and their engineers worked closely with us to develop the necessary tooling and infrastructure. All we had to do was communicate our needs with Scale. This greatly freed up our engineering time and resources to tackle other urgent problems while still acquiring the critical data we needed,” said Wang. Scale was also able to cover a wide distribution of questions to provide a greater lift in model performance.
Cohere's Command LLM Improves with High-Quality Data
The data provided by Scale accounted for a sizable portion of Cohere’s first Command dataset. By collaborating on instruction-following best practices and doubling down on complex, tightly paired data, Scale contributed to performance gains for Cohere’s Command model. Command is one of the highest-performing LLMs across HELM and other standardized metrics. Cohere retrains its LLMs on a weekly basis to constantly improve the models, and that requires a regular stream of high-quality data.
“Scale has been incredibly responsive to our data needs, both at a macro level (designing novel workflows for collecting complex types of data) and a micro level (adjusting existing data collection protocols to better align the data with our needs). We give feedback to Scale weekly, and they are always receptive to our feedback, positive or negative. If we request a change, they are quick to move and have the change implemented within a couple of weeks. We feel like valued customers at Scale,” concluded Wang.