TrafficInstruct - MDD Fine Macro F1: leaderboard
TrafficInstruct v1; released Qwen3.8-27B without task-specific tuning/LoRA. Fixed train/validation/test splits; known task input, deterministic serialized traffic, structured-answer parser. Coarse predictions project fine labels; fine metrics use full test set and fixed fine vocabulary. Invalid/missing/unknown primary labels are incorrect; MacroF1 uses all fixed labels. Plaintext/encrypted Web conditions and each task kept separate. Task-trained LoRA checkpoints and non-language RF/ET-BERT comparators not aliased to the released model.
Metric: Macro F1 (%). Source: arxiv.org. 1 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 27B | 3.8 | #123 |
Interactive version: theaggregate.ai/benchmark?slug=trafficinstruct-mdd-fine-macro-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.