Mineral Identification from Field Observations

ranked by score ↓
Source

Paste as source: in your trap.yaml

git+https://github.com/Zhuaiz/elastic-mineral-hackson@dfa77e12bf7db981465d9b54734b985ecd14ec99#subdirectory=trap/task
Share

mineral-species-id

Given a field geologist's hand-specimen observations of an unknown mineral — crystal system, Mohs hardness, streak color, body color, luster, and specific gravity — name the single most likely mineral species out of 98 candidates. This task measures how well a model identifies mi

Task

Each case gives observed physical properties (never the chemical formula — that would give the answer away) and asks for the species name only, one lowercase word. The reference document is a 98-species property handbook so retrieval-augmented runs can match observations against catalog entries; closed-book runs must recall the species from the properties alone.

Scoring

The judge normalizes the answer and reference species name (lowercasing, stripping punctuation, and collapsing known synonyms and transliterations such as creedite/credit, stibnite/antimonite, labradorite/labrador) and requires an exact match. One point per correct species, accuracy over all cases.

Data and licensing

Reference descriptions are excerpts from English Wikipedia (CC BY-SA 4.0). Physical-property values (Mohs hardness, streak, crystal system) are objective measurements. No Mindat data or specimen images are redistributed here.