Models and benchmarks
Models and benchmarks — the building blocks of the Stratix catalog.
Last updated
Was this helpful?
⌘Ctrlk
Models and benchmarks — the building blocks of the Stratix catalog.
The two primitives that anchor every Stratix evaluation.
A specific LLM identified by provider/model-name (e.g., openai/gpt-4o, anthropic/claude-opus-4-7). In Stratix:
Public models live in the global catalog (200+)
Private models are BYOK custom models in your org
Each model has metadata: provider, capabilities, context window, modalities, licensing, pricing, score history.
The model is the substrate. Picking the wrong substrate is the most expensive AI mistake.
A standardized dataset and grading procedure. In Stratix:
Public benchmarks live in the catalog (52+, e.g., MMLU, HumanEval, GSM8K)
Private benchmarks are datasets you upload to your org
Each benchmark has metadata: capability tag, sample size, methodology, license, score history.
Benchmarks are how you make "is the model good?" measurable. Pick benchmarks whose distribution matches your task; ignore the ones that don't.
Every public evaluation is (model, benchmark, scoring config). The leaderboard is "all public evaluations of benchmark X, sorted by score."
Both models and benchmarks carry capability tags:
chat — general dialog
code — code generation
reasoning — chain-of-thought, math, logic
multilingual
vision
multi-turn
tool-use
When picking, match the model's strongest tag against the benchmark's tag.
The catalog preserves history. A model evaluated against MMLU last quarter gets a score; if it's re-evaluated against an updated MMLU, both scores remain visible with methodology versions.
A benchmark is not a "test set" in the ML sense. It's a public, curated, citable evaluation harness. Your private dataset is your "test set."
A model's score on benchmark X doesn't predict its score on your data. Public scores are a leading indicator; your private evaluation is the verdict.
Last updated
Was this helpful?
Was this helpful?