From Perplexity · · 3m
Perplexity Decider v1.3: Training, eval, and use cases
Perplexity releases Decider v1.3 and Flash v1.3 decision models with improved accuracy and new locate feature for text and vision tasks.
In brief
Perplexity releases Decider v1.3 and Flash v1.3 decision models with improved accuracy and new locate feature for text and vision tasks. Perplexity launches Decider v1.3 and Flash v1.3 decision models that outperform benchmarks in classification, reranking, and entity extraction tasks. Originally reported by…
Quick answers
- What is a decision model?
- A decision model is a general-purpose classifier that returns probabilities for multiple questions about an input in a single forward pass.
- What is the key takeaway from Perplexity Decider v1.3: Training, eval, and use cases?
- Perplexity launches Decider v1.3 and Flash v1.3 decision models that outperform benchmarks in classification, reranking, and entity extraction tasks.
What is a decision model?
A decision model is a general-purpose classifier that returns probabilities for multiple questions about an input in a single forward pass.
Decider v1.3 features
Decider v1.3 and Flash v1.3 models improve performance and robustness across text and vision tasks through bidirectional attention, better-curated training data, and refined training recipes.
New locate feature
Locate extracts text from the state for named entity recognition, personal information identification, and extractive question-answering tasks without changing ordinary request handling.
Training process
Decider is trained using three stages: generalist training, fine-grained task-specific tuning with LoRA, and vision-specific specialization to balance text and visual performance.
Bidirectional attention
Models use bidirectional attention to let each token incorporate context from the entire input, improving performance and stability over causal attention.
Locate training
Locate is trained with an adapted InfoNCE loss to push correct spans above a threshold while keeping wrong spans below, using a span head and LoRA adapters.
Evaluation benchmarks
Decider v1.3 outperforms Jev 1.14 across 200+ internal tasks, multilingual classification, and vision tasks with 76.99% accuracy on five visual benchmarks.
Real-world performance
In production, Decider v1.3 achieves 93.98% agreement with LLM judges across 2.3 million user queries, outperforming previous versions by 0.34 percentage points.
Self-improving security
Decider identifies flaws in security rules, leading to 14 rule fixes in one day with 6 approved within two days, cutting rule volume by 52%.
Monitoring RL rollouts
Decider monitors reinforcement learning rollouts by answering multiple-choice questions about failed tool calls and incorrect interpretations.
Page quality filtering
Decider filters low-quality web pages by assessing content quality, domain, audience, and harmful information in real-time during crawling.
Visual QA capabilities
Decider reviews style changes to web pages or apps, identifying clipped text, overlapping elements, broken layouts, and misleading charts.
Use cases
Decider supports planning research tasks, browser navigation, and visual QA by estimating complexity before search begins and providing detailed state information.
Model selection
Choose Decider v1.3 for nuanced judgments and Decider Flash v1.3 for high-volume routing where latency and cost are critical.
Availability
Both models are available on Hugging Face under Apache license, through OpenRouter, and the Perplexity Decisions API.
What matters most
Perplexity launches Decider v1.3 and Flash v1.3 decision models that outperform benchmarks in classification, reranking, and entity extraction tasks.
