Thank you for sending your enquiry! One of our team members will contact you shortly.
Thank you for sending your booking! One of our team members will contact you shortly.
Duration 7 hours
Course Outline
Transitioning from Text Generation to Typed Decisions
- Nature of outputs returned by decision models
- Contrasting system one decisions with reasoning LLMs
- Cost and latency characteristics of bounded decisions
The Ecosystem of Open-Weight Decision Models
- AWS Labs' Strands Decider 2B
- Open-weight families including JEV and GEV
- Models such as Pplx-Decider, Liquid D1, and Clef
- Variances in licensing and deployment requirements
Architectural Insights into Decision Models
- Utilizing an LLM torso without a standard language modelling head
- Scoring candidate options via the pointer head mechanism
- Origins of the confidence score calculation
Executing Initial Decision Tasks
- Local configuration on CPU and consumer-grade GPUs
- Formulating effective questions and option sets
- Processing multiple questions in a single inference pass
- Managing thresholds and defining fallback behaviors
Evaluating Decision Quality
- Utilizing JevBench and other decision-specific benchmarks
- Balancing accuracy against calibration performance
- Interpreting Brier scores and the reliability of confidence levels
- Identifying tasks where decision models are ill-suited
Requirements
- Fundamental understanding of large language models and prompt-based applications.
- Proficiency with command-line interfaces and Python scripting.
- No prior experience in machine learning training is necessary.
Target Audience
- AI engineers assessing decision models for agent-centric workflows.
- Software developers looking to substitute unbounded LLM calls with constrained decisions.
- Solution and platform architects focused on designing agent architectures.
- Technical product managers evaluating decision model capabilities.
