Adaptive Engine is Adaptive ML's platform for tuning, evaluating, and serving language models with reinforcement learning. It is aimed at teams that want to specialise open models on their own tasks and keep the resulting model, using a measurable objective as the tuning signal.
Documented deployments describe the pattern. Aïkan built Juribot, a chatbot answering questions about insurance documents, where off-the-shelf models can cite laws that do not exist; a Llama 3.1 8B model tuned on the platform with synthetic annotations reduced hallucinations by 25% against GPT-4o and by 42% against the base model. SK Telecom tuned open models as small as Gemma 3 4B for multilingual content moderation. AT&T deployed the platform as its reinforcement-tuning system and identified more than 50 candidate use cases spanning text-to-SQL, customer support, call summarisation, and document retrieval-augmented generation; during its evaluation a tuned Llama 3.1 8B reached a 51% win rate against GPT-4o on retrieval over telecoms documents.
The company behind the platform has been acquired by Datadog.
Features
- Reinforcement tuning: models are trained against a measurable objective rather than a fixed supervised dataset
- Evaluation: candidate models are compared on the task metric before and after tuning
- Serving: tuned models are deployed from the same platform
- Small open models: documented deployments tune models in the 4B to 8B range, including Llama 3.1 8B and Gemma 3 4B
- Synthetic annotations: training data can be generated rather than hand-labelled, which shortens production time
