Perplexity Decider v1.3: Training, eval, and use cases

Perplexity Decider v1.3 and its smaller distilled version, Decider Flash v1.3, are general-purpose decision models for classification and scoring rather than free-form text generation. A request supplies shared context and one or more questions; the model returns probabilities across yes/no, custom multiple-choice, or ordered score labels in a single pass. This makes the models suitable for judging output quality and safety, routing requests, moderation, tagging, and ranking search results. The release also adds LOCATE, which identifies requested text spans - such as names, personal information, or answers to extractive questions - and returns their exact character positions. Both models are offered under the Apache license on Hugging Face, through OpenRouter, and via Perplexity’s Decisions API.

The models adapt generative LLMs using bidirectional attention and a three-stage training process: broad fine-tuning, targeted LoRA specialization, and vision-focused specialization. The larger model is trained first, then teaches the smaller model through probability-distribution distillation, which Perplexity says generalizes better to held-out data than training directly on labels. The chosen “isolated questions” attention pattern lets questions share context while reducing the chance that one question changes another’s answer. LOCATE uses a separate span-prediction head and adapters, kept unmerged because combining them with the main model still causes some regression in decision performance. The release reports improved robustness across text and vision tasks; its evaluation draws on external benchmarks and includes steps to reduce dataset overlap and bias, though the supplied description does not give benchmark results.