Tarsier2-7B: benchmark results
Provider: Other. Released 2025-04-25. Access: Open.
Aggregate Rating 1618 ± 38, rank #854 of 10916 rated models, from 27 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MoHallBench - Semantic Similarity Hallucination (Multiple Choice) | 87.5 | Multiple-choice accuracy (%; four-option multiple choice (gr | 100 |
| MoHallBench - Sequential Inference Hallucination | 29.1 | Q-PairAcc (%; share of adversarial pairs in which the model | 100 |
| MoHallBench - Sequential Inference Hallucination Rate (Generative) | 5.7 | Hallucination rate (%; share of free-form descriptions of tr | 100 |
| MoHallBench - Action-Action Hallucination (Multiple Choice) | 80.8 | Multiple-choice accuracy (%; four-option multiple choice (gr | 94.4 |
| Reviso Captioning - DREAM-1K Precision | 41.5 | Precision (%) | 93.3 |
| MoHallBench - Action-Action Hallucination | 26.9 | Q-PairAcc (%; share of adversarial pairs in which the model | 88.9 |
| MoHallBench - Object-Action Hallucination (Multiple Choice) | 73.7 | Multiple-choice accuracy (%; four-option multiple choice (gr | 88.9 |
| MoHallBench - Scene-Action Hallucination (Multiple Choice) | 82.4 | Multiple-choice accuracy (%; four-option multiple choice (gr | 88.9 |
| Reviso Captioning - DREAM-1K F1 | 39.2 | F1 (%) | 86.7 |
| MoHallBench - Semantic Similarity Hallucination | 27.6 | Q-PairAcc (%; share of adversarial pairs in which the model | 77.8 |
| MoHallBench - Visual Similarity Hallucination | 31.7 | Q-PairAcc (%; share of adversarial pairs in which the model | 77.8 |
| TVBench | 54.7 | Average Accuracy | 74.1 |
Interactive version: theaggregate.ai/model?slug=tarsier2-7b · How It Works · Data refreshed daily, snapshot 2026-11-08.