Few shot triplet
Three worked examples covering the flat, nested, and missing field cases. Costs more tokens than any other entry in this suite.
Prompt
// system
Extract JSON from the input. Follow the pattern of these examples exactly.
// user
Examples:
{{examples}}
Now extract from: {{document}}
Score history
Model versionVersionScore · CIHoldoutState
claude-fable-5v184.5 78.3–89.283.3fresh
llama-4.2-405bv169.6 62.3–76.169.4fresh
gpt-5.6-solv183.3 77.0–88.283.3fresh
claude-opus-4.8v181.5 75.0–86.780.6fresh
mistral-large-3v174.4 67.3–80.473.6fresh
gemini-3.2-prov179.2 72.4–84.677.8fresh
current scorev1
84.5
95% CI 78.3 to 89.2 · n=168
Holdout score83.3
Judgeprogrammatic
Cost per run$1.850
Median latency2.7s
LicenseCC-BY-4.0
Author@nils
0 endorsements
Sign in to endorse this entry.
Across models
claude-fable-584.5fresh
llama-4.2-405b69.6fresh
gpt-5.6-sol83.3fresh
claude-opus-4.881.5fresh
mistral-large-374.4fresh
gemini-3.2-pro79.2fresh
Versions
v1First submission84.5