Datasets:
Pocket-Dentist-Bench
🚧 Dataset Coming Soon — The full benchmark data will be released upon paper acceptance.
Please check back later or ⭐ star this repo to get notified.
Links
- 📄 Paper: Coming soon
- 💻 Code: GitHub
- 🤗 Models: Coming soon
Overview
Pocket-Dentist is a large-scale multimodal benchmark and deployment pipeline for evaluating Vision-Language Models (VLMs) on dental image understanding tasks. The benchmark curates and standardizes seven dental datasets into a unified vision-language evaluation framework.
Key Statistics:
- 🏥 6,000+ patients
- 🖼️ 71,000+ images
- 📷 4 imaging modalities
- 📋 6 task types
- 📊 14 evaluation metrics
- 🤖 14 VLMs benchmarked (including 12 open-weight models with LoRA adaptation)
Benchmark Pipeline
The Pocket-Dentist evaluation pipeline consists of four stages:
- Data Collection & Unification — Curating heterogeneous dental datasets into a unified multimodal benchmark
- Task Design & Annotation — Converting source annotations into shared prompt–response task formats
- Model Evaluation & Adaptation — Evaluating VLMs under zero-shot, few-shot, and LoRA adaptation settings
- On-Device Deployment — Measuring local inference efficiency for compact adapted models on mobile hardware
Figure 1: Deploy-aware evaluation pipeline of Pocket-Dentist.
Benchmark Results
Zero-Shot Performance
Under zero-shot evaluation, no single model dominates across all 14 metrics. Closed-source APIs (Gemini) perform best overall, while compact model performance is fragmented across tasks.
Table 1: Zero-Shot (ZS) Results — Click to expand
Large VLMs (≥7B)
| Model | BRAR Acc | BRAR F1 | DR F1w | Meta VQA | Meta Cap | Meta Cls | Aariz VQA | Aariz CVM | COde Cls | DenPAR Arch | DenPAR Site | DenPAR MAE↓ | Caries Det | Caries Cls |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Lingshu-32B | 0.49 | 0.39 | 0.60 | 0.63 | 0.18 | 0.34 | 0.26 | 0.13 | 0.48 | 0.59 | 0.53 | 0.88 | 0.56 | 0.16 |
| MedMO-8B-Next | 0.26 | 0.19 | 0.53 | 0.49 | 0.09 | 0.08 | 0.21 | 0.05 | 0.26 | 0.61 | 0.29 | 2.90 | 0.59 | 0.84 |
| Qwen2.5-VL-7B | 0.27 | 0.17 | 0.32 | 0.45 | 0.15 | 0.23 | 0.20 | 0.00 | 0.50 | 0.40 | 0.35 | 1.01 | 0.63 | 0.14 |
| gemini-2.0-flash | 0.57 | 0.37 | 0.00 | 0.63 | 0.18 | 0.36 | 0.29 | 0.25 | 0.54 | 0.84 | 0.45 | 0.42 | 0.50 | 0.12 |
| gemini-2.5-flash | 0.27 | 0.26 | 0.62 | 0.66 | 0.14 | 0.24 | 0.23 | 0.12 | 0.58 | 0.99 | 0.51 | 0.47 | 0.54 | 0.13 |
Compact VLMs (≤4B)
| Model | BRAR Acc | BRAR F1 | DR F1w | Meta VQA | Meta Cap | Meta Cls | Aariz VQA | Aariz CVM | COde Cls | DenPAR Arch | DenPAR Site | DenPAR MAE↓ | Caries Det | Caries Cls |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3.5-4B | 0.17 | 0.10 | 0.54 | 0.82 | 0.10 | 0.16 | 0.17 | 0.04 | 0.11 | 0.40 | 0.19 | 3.02 | 0.49 | 0.18 |
| Qwen3-VL-4B | 0.44 | 0.37 | 0.24 | 0.58 | 0.20 | 0.22 | 0.23 | 0.08 | 0.54 | 0.44 | 0.23 | 0.42 | 0.63 | 0.58 |
| gemma-4-E4B-it | 0.56 | 0.24 | 0.61 | 0.59 | 0.18 | 0.31 | 0.31 | 0.04 | 0.51 | 0.40 | 0.51 | 0.52 | 0.43 | 0.30 |
| medgemma-4b-it | 0.44 | 0.33 | 0.57 | 0.54 | 0.14 | 0.16 | 0.40 | 0.03 | 0.27 | 0.40 | 0.23 | 0.89 | 0.52 | 0.11 |
| paligemma2-3b | 0.10 | 0.06 | 0.00 | 0.00 | 0.00 | 0.00 | 0.20 | 0.03 | 0.00 | 0.00 | 0.18 | 0.89 | 0.64 | 0.00 |
| SmolVLM2-2.2B | 0.56 | 0.35 | 0.56 | 0.00 | 0.10 | 0.15 | 0.23 | 0.05 | 0.10 | 0.60 | 0.10 | 0.89 | 0.44 | 0.92 |
| InternVL3.5-2B | 0.50 | 0.27 | 0.09 | 0.15 | 0.00 | 0.00 | 0.37 | 0.00 | 0.14 | 0.40 | 0.22 | 3.21 | 0.36 | 0.12 |
| gemma-4-E2B-it | 0.56 | 0.24 | 0.24 | 0.48 | 0.15 | 0.25 | 0.39 | 0.03 | 0.50 | 0.27 | 0.23 | 0.73 | 0.61 | 0.11 |
| InternVL3.5-1B | 0.26 | 0.14 | 0.62 | 0.34 | 0.00 | 0.00 | 0.21 | 0.07 | 0.14 | 0.28 | 0.19 | 2.27 | 0.61 | 0.11 |
Bold = best in tier. ↑ higher is better; MAE ↓ lower is better.
On-Device Deployment
We deploy LoRA-tuned VLMs on an iPhone 17 Pro (A19 Pro SoC, 12 GB Unified Memory) via Metal-accelerated inference using llama.cpp. All computation is performed locally on the device with 100% offline privacy protection.
| Model | Total Latency (s) ↓ | TTFT (s) ↓ | ITPS (t/s) ↑ | OTPS (t/s) ↑ | RAM (GB) ↓ |
|---|---|---|---|---|---|
| Pocket-Dentist-4B | 6.67 | 1.22 | 315.95 | 17.07 | 4.09 |
| InternVL3.5-2B | 4.74 | 0.78 | 434.53 | 29.47 | 2.62 |
| Qwen2.5-VL-7B | 24.06 | 2.29 | 148.93 | 9.60 | 6.22 |
Pocket-Dentist iOS app running Pocket-Dentist-4B locally on an iPhone 17 Pro.
Key Findings
- 🔍 Zero-shot fragmentation: No single model dominates across all dental tasks under zero-shot evaluation
- 📈 LoRA adaptation closes the gap: Under a uniform LoRA budget, compact VLMs become competitive with substantially larger models
- 🏆 Qwen3-VL-4B achieves the strongest overall performance among compact models, matching or outperforming larger open-weight models (7B–32B) on most primary task metrics
- 📱 Pocket-Dentist-4B (LoRA-tuned Qwen3-VL-4B) runs locally on an iPhone 17 Pro with 6.67s per-sample latency and 4.09 GB RAM
- 🏥 Medical pre-training alone does not guarantee dental task performance — dental-domain LoRA adaptation is more effective
License & Disclaimer
This benchmark integrates data from multiple publicly available dental imaging datasets, each with its own license. This repository is distributed under CC BY-NC-SA 4.0. Users must also comply with the individual licenses of the constituent datasets.
- Downloads last month
- 14