This model is a fine-tuned version of the Qwen3-VL-8B-Instruct model using the Unsloth library on the MBZUAI/TimeTravel dataset.
Model Details
Base Model: Qwen3-VL-8B-Instruct-unsloth-bnb-4bit
Fine-tuning Library: Unsloth
Dataset: MBZUAI/TimeTravel
Task: Image description generation for historical artifacts.
Training
The model was trained for 4 epochs with a batch size of 8 (4 per device with 8 gradient accumulation steps). LoRA adapters were used for parameter efficient finetuning, targeting both vision and language layers, as well as attention and MLP modules.
Dataset
TimeTravel: A Comprehensive Benchmark to Evaluate LMMs on Historical and Cultural Artifacts
TimeTravel Taxonomy maps artifacts from 10 civilizations, 266 cultures, and 10k+ verified samples for AI-driven historical analysis.
tax
🌟 Key Features
TimeTravel is the first large-scale, open-source benchmark designed to evaluate Large Multimodal Models (LMMs) on historical and cultural artifacts. It covers:
266 Cultural Groups across 10 Historical Regions
10,000+ Expert-Verified Artifact Samples
Multimodal Image-Text Dataset for AI-driven historical research
A publicly available dataset and evaluation framework to advance AI applications in history and archaeology.
🔄 TimeTravel Creation Pipeline
The TimeTravel dataset follows a structured pipeline to ensure the accuracy, completeness, and contextual richness of historical artifacts.
pipeline
Our approach consists of four key phases:
Data Selection: Curated 10,250 artifacts from museum collections, spanning 266 cultural groups, with expert validation to ensure historical accuracy and diversity.
Data Cleaning: Addressed missing or incomplete metadata (titles, dates, iconography) by cross-referencing museum archives and academic sources, ensuring data consistency.
Generation & Verification: Used GPT-4o to generate context-aware descriptions, which were refined and validated by historians and archaeologists for authenticity.
Data Aggregation: Standardized and structured dataset into image-text pairs, making it a valuable resource for AI-driven historical analysis and cultural heritage research.
🏆 TimeTravel Evaluation
The table below showcases the performance comparison of various closed and open-source models on our proposed TimeTravel benchmark.
Model
BLEU
METEOR
ROUGE-L
SPICE
BERTScore
LLM-Judge
GPT-4o-0806
0.1758🏅
0.2439
0.1230🏅
0.1035🏅
0.8349🏅
0.3013🏅
Gemini-2.0-Flash
0.1072
0.2456
0.0884
0.0919
0.8127
0.2630
Gemini-1.5-Pro
0.1067
0.2406
0.0848
0.0901
0.8172
0.2276
GPT-4o-mini-0718
0.1369
0.2658🏅
0.1027
0.1001
0.8283
0.2492
Llama-3.2-Vision-Inst
0.1161
0.2072
0.1027
0.0648
0.8111
0.1255
Qwen-2.5-VL
0.1155
0.2648
0.0887
0.1002
0.8198
0.1792
Llava-Next
0.1118
0.2340
0.0961
0.0799
0.8246
0.1161
Model
India
Roman Emp.
China
British Isles
Iran
Iraq
Japan
Cent. America
Greece
Egypt
GPT-4o-0806
0.2491🏅
0.4463🏅
0.2491🏅
0.1899🏅
0.3522🏅
0.3545🏅
0.2228🏅
0.3144🏅
0.2757🏅
0.3649🏅
Gemini-2.0-Flash
0.1859
0.3358
0.2059
0.1556
0.3376
0.3071
0.2000
0.2677
0.2582
0.3602
Gemini-1.5-Pro
0.1118
0.2632
0.2139
0.1545
0.3320
0.2587
0.1871
0.2708
0.2088
0.2908
GPT-4o-mini-0718
0.2311
0.3612
0.2207
0.1866
0.2991
0.2632
0.2087
0.3195
0.2101
0.2501
Llama-3.2-Vision-Inst
0.0744
0.1450
0.1227
0.0777
0.2000
0.1155
0.1075
0.1553
0.1351
0.1201
Qwen-2.5-VL
0.0888
0.1578
0.1192
0.1713
0.2515
0.1576
0.1771
0.1442
0.1442
0.2660
Llava-Next
0.0788
0.0961
0.1455
0.1091
0.1464
0.1194
0.1353
0.1917
0.1111
0.0709
🖼 TimeTravel Examples
The figure illustrates the cultural and material diversity of the TimeTravel dataset.
tax
📂 TimeTravle Dataset Schema
📷 Image (image)
🔹 id (string)
📅 Production date (string)
📍 Find spot (string)
🔸 Materials (string)
🛠 Technique (string)
📝 Inscription (string)
🎭 Subjects (string)
📛 Assoc name (string)
🏛 Culture (string)
📂 Section (string)
🌍 Place (string)
📝 description (string)
📚 Citation
If you use TimeTravle dataset in your research, please consider citing:
bibtex
1@misc{ghaboura2025timetravelcomprehensivebenchmark,
2 title={Time Travel: A Comprehensive Benchmark to Evaluate LMMs on Historical and Cultural Artifacts},
3 author={Sara Ghaboura and Ketan More and Ritesh Thawkar and Wafa Alghallabi and Omkar Thawakar and Fahad Shahbaz Khan and Hisham Cholakkal and Salman Khan and Rao Muhammad Anwer},
4 year={2025},
5 eprint={2502.14865},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/2502.14865},
9}
Usage
This model can be used for generating descriptions of historical artifact images.
