MADE - Tail Labels: leaderboard

Rare-label classification on MADE (2026), a living benchmark of FDA medical device adverse event reports: macro-F1 over the 348 tail labels attached to 0.01 to 0.1 percent of training reports, with models prompted with ten kNN-retrieved training reports. Percent; higher is better.

Metric: Macro-F1 (%, the 0-1 score times 100) over the 348 tail labels (0.01 to 0.1 percent of training reports) with ten-shot prompting with kNN-retrieved training reports, the task instructions and the label list, on the 10,288-report stratified (truncated) MADE test set of FDA medical device adverse event reports from July 2024 to June 2025 with 1,154 hierarchical IMDRF product and patient problem labels, greedy decoding (GPT-5 at its default medium reasoning effort); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-553#178
2Qwen 3 235B A22BThinking48#628
3DeepSeek R1 052847#428
4Qwen 3 30B A3BThinking44#1156
5GPT-4.142#505
6Qwen 3 235B A22B Instruct42#691
7GLM 4.5 Air39#886
8GPT-OSS-120B38#704
9Llama-3.3-nemotron-super-49B-v1.538#920
10Qwen 3 4BReasoning36#2348
11Llama 3.1 70B Instruct25#1571
12Qwen 3 4B Instruct25#1696
13Qwen 3 30B A3B Instruct14#1053
14Kimi K26#443
15Llama 3.1 8B Instruct3#4600

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=made-tail-labels · How It Works · Data refreshed daily, snapshot 2026-11-21.