What does professional blending work actually look like, step by step? This sample dataset holds 294 annotated steps across 14 blending-bench trajectories, recorded as working tasters built real red blends and single-vineyard lots at the bench and in the barrel cellar.
We gave five tasters from the Oenra network a single cellar brief each, a new-release red blend, and recorded them carrying it through to multiple bottling-ready deliverables: a bench blend, a barrel selection, a press-fraction trial, a final assembly. The tasters narrated their reasoning aloud as they worked, so every step's thought is transcribed from the expert's own voice.
Each step pairs a photograph with a first-person thought, a structured action, and reproducible grounding: a component poured, a blend ratio set, a descriptor called, or a score logged, with a preferred execution path marking the most deterministic option. The format follows the AgentNet trajectory schema, extended with a blending action taxonomy and multi-path execution. This structure makes the data directly usable in several contexts:
- Palate preference modeling: long-horizon, multi-cellar trajectories with photographs, actions, and blend ratios in a format interoperable with existing sensory-panel tooling.
- Taster calibration: pair photographs with narration-grounded reasoning to capture how experts actually plan, taste, and correct course mid-assembly.
- Sensory scoring systems: every step carries a structured action, so a bench can reach the same blend through a shared rubric instead of loose prose.
- Benchmarking and evals: a human expert baseline for long-horizon cellar work, reproducible end-to-end at the bench.
The dataset was recorded in 2026 at the blending bench and in the barrel cellar. It is released under the CC BY 4.0 license on Hugging Face.