FlagEval is an evaluation toolkit for AI large foundation models. Our goal is to explore and integrate scientific, fair and open foundation model evaluation benchmarks, methods and tools. FlagEval will support multi-dimensional evaluation (such as accuracy, efficiency, robustness, etc.) of foundation models in/cross different modalities (such as NLP, audio, CV and multimodal) in the future. We hope that through the evaluation of the foundation models, we can deepen the understanding of the foundation models and promote related technological innovation and industrial application.
mCLIPEval is a evaluation toolkit for vision-language models (such as CLIP, Contrastive Language–Image Pre-training).
- Including Multilingual (12 languages) datasets and monolingual (English/Chinese) datasets.
- Supporting for Zero-shot classification, Zero-shot retrieval and zeroshot composition tasks.
- Adapted to FlagAI pretrained models (AltCLIP, EVA-CLIP), OpenCLIP pretrained models, Chinese CLIP models, Multilingual CLIP models, Taiyi Series pretrained models, or customized models.
- Data preparation from various resources, like torchvision, huggingface, kaggle, etc.
- Visualization of evaluation results through leaderboard figures or tables, and detailed comparsions between two specific models.
Environment Preparation:
- Pytorch version >= 1.8.0
- Python version >= 3.8
- For evaluating models on GPUs, you'll also need install CUDA and NCCL
Step:
git clone https://github.com/FlagOpen/FlagEval.git
cd FlagEval/mCLIPEval/
pip install -r requirements.txtPlease refer to mCLIPEval/README.md for more details.
ImageEval-prompt is a set of prompts that evaluate text-to-image (T2I) models at a fine-grained level, including entity, style and detail. By conducting comprehensive evaluations at a fine-grained level, researchers can better understand the strengths and limitations of T2I models, in order to further improve their performance.
- Including 1,624 English prompts and 339 Chinese prompts.
- Each prompt is annotated using "double-blind annotation & third-party arbitration" approach, divided into three dimensions: entities, styles, and details.
- Entity dimension includes five sub-dimensions: object, state, color, quantity, and position;
- Style dimension includes two sub-dimensions: painting style and cultural style;
- Detail dimension includes four sub-dimensions: hands, facial features, gender, and illogical knowledge.
Please refer to imageEval/README.md for more details.
- For help and issues associated with FlagEval, or reporting a bug, please open a GitHub Issue or e-mail to flageval@baai.ac.cn. Let's build a better & stronger FlagEval together :)
- We're hiring! If you are interested in working with us on foundation model evaluation, please contact flageval@baai.ac.cn.
- Welcome to collaborate with FlagEval! New task or new dataset submissions are encouraged. If you are interested in contributiong new task or new dataset or new tool to FlagEval, please contact flageval@baai.ac.cn.
The majority of FlagEval is licensed under the Apache 2.0 license, however portions of the project are available under separate license terms:
- The usage of CLIP_benchmark is licensed under the MIT license
- The usage of ImageNet1k datasets in under the huggingface datasets license and ImageNet licenese