--- title: CodeFuse-VLM-14B canonical_url: "https://www.modelscope.cn/models/codefuse-ai/CodeFuse-VLM-14B" md_url: "https://www.modelscope.cn/models/codefuse-ai/CodeFuse-VLM-14B.md" repository: codefuse-ai/CodeFuse-VLM-14B last_updated: 2024-04-18 pipeline_tag: multimodal-dialogue tasks: - multimodal-dialogue model_type: - qwen architectures: - QWenLMHeadModel library_name: - pytorch frameworks: - Pytorch inference_backends: - "deploy_task text/emb" - "lmdeploy 0.9.1" - "lmdeploy_turbomind 0.9.1" - "sglang 0.5.2" - "vllm 0.9.2" downloads: 303 stars: 4 --- # CodeFuse-VLM-14B > CodeFuse-VLM-14B - codefuse-ai 在 ModelScope 开源的模型。CodeFuse-VLM is a Multimodal LLM(MLLM) framework that provides users with multiple vision encoders, multimodal alignment adapters, and LLMs. codefuse-ai/CodeFuse-VLM-14B 是 ModelScope 魔搭社区上的multimodal-dialogue模型，可用 deploy_task text/emb、lmdeploy 0.9.1、lmdeploy_turbomind 0.9.1 部署。 - **Repository**: codefuse-ai/CodeFuse-VLM-14B - **Tasks**: multimodal-dialogue - **Inference backends**: deploy_task text/emb, lmdeploy 0.9.1, lmdeploy_turbomind 0.9.1, sglang 0.5.2, vllm 0.9.2 - **Downloads**: 303 - **Stars**: 4 - **Last updated**: 2024-04-18 Source: https://www.modelscope.cn/models/codefuse-ai/CodeFuse-VLM-14B ---

## CodeFuse-VLM CodeFuse-VLM is a Multimodal LLM(MLLM) framework that provides users with multiple vision encoders, multimodal alignment adapters, and LLMs. Through CodeFuse-VLM framework, users are able to customize their own MLLM model to adapt their own tasks. As more and more models are published on Huggingface community, there will be more open-source vision encoders and LLMs. Each of these models has their own specialties, e.g. Code-LLama is good at code-related tasks but has poor performance for Chinese tasks. Therefore, we built CodeFuse-VLM framework to support multiple vision encoders, multimodal alignment adapters, and LLMs to adapt different types of tasks.

Under CodeFuse-VLM framework, we use cross attention multimodal adapter, Qwen-14B LLM, and Qwen-VL's vision encoder to train CodeFuse-VLM-14B model. On multiple benchmarks, our CodeFuse-VLM-14B shows superior performances over Qwen-VL and LLAVA-1.5.

Here is the table for different MLLM model's performance on benchmarks Model | MMBench | MMBench-CN | VqaV2 | GQA | TextVQA | Vizwiz | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | LLAVA-1.5 | 67.7 | 63.6 | 80.0 | 63.3 | 61.3 | 53.6 Qwen-VL | 60.6 | 56.7 | 78.2 | 57.5 | 63.8 | 38.9 CodeFuse-VLM-14B | 75.7 | 69.8 | 79.3 | 59.4 | 63.9 | 45.3 ## Contents - [Install](#Install) - [Datasets](#Datasets) - [Multimodal Alignment](#Multimodal-Alignment) - [Visual Instruction Tuning](#Visual-Instruction-Tuning) - [Evaluation](#Evaluation) ## Install Please run sh init\_env.sh ## Datasets Here's the table of datasets we used to train CodeFuse-VLM-14B: Dataset | Task Type | Number of Samples | ------------- | ------------- | ------------- | synthdog-en | OCR | 800,000 synthdog-zh | OCR | 800,000 cc3m(downsampled)| Image Caption | 600,000 cc3m(downsampled)| Image Caption | 600,000 SBU | Image Caption | 850,000 Visual Genome VQA (Downsampled) | Visual Question Answer(VQA) | 500,000 Visual Genome Region descriptions (Downsampled) | Reference Grouding | 500,000 Visual Genome objects (Downsampled) | Grounded Caption | 500,000 OCR VQA (Downsampled) | OCR and VQA | 500,000 Please download these datasets on their own official websites. ## Multimodal Alignment Please run sh scripts/pretrain.sh or sh scripts/pretrain\_multinode.sh ## Visual Instruction Tuning Please run sh scripts/finetune.sh or sh scripts/finetune\_multinode.sh ## Evaluation Please run python scripts in directory llava/eval/. Our pre-trained CodeFuse-VLM-14B can be loaded with the following code: ``` import os from llava.model.builder import load_mixed_pretrained_model model_path = '/pretrained/model/path' tokenizer, model, image_processor, context_len = load_mixed_pretrained_model(model_path, None, 'qwen-vl-14b', os.path.join(model_path, 'Qwen-VL-visual'), 'cross_attn', os.path.join(model_path, 'mm_projector/mm_projector.bin')) ``` You can also run scripts/merge\_qwen\_vl\_weights.sh first and load the merged model by the following code: ``` from llava.model import LlavaQWenForCausalLM model = LlavaQWenForCausalLM.from_pretrained('/path/to/our/pretrained/model') ``` ## Join Us We are the **AI Native** team within the Platform Technology Business Group at Ant Group, dedicated to the intelligentization of Ant Group's platform engineering. Established for over three years, our team has played a pivotal role in supporting the intelligent operation and maintenance of Ant Group's cloud computing infrastructure. Our mission is to build algorithm services and platforms with a wide user base through world-class technological innovation and impact, supporting the implementation of internal and external products and businesses. Embracing an innovation-driven ethos, our team not only supports business implementation but also propels technological influence. Over the past three years, we have published more than 20 papers at top conferences like ICLR, NeurIPS, KDD, and ACL. Our innovative business outcomes have earned us two Ant Technology's highest T-Star awards and one SuperMA award from Ant Group. Our open-source project CodeFuse has received 4K stars as of February 2024, and our models have been downloaded over 1.5 million times on Huggingface and Modelscope. **We are on the lookout for top talents to join our vibrant team! If you're eager to develop your career in an environment filled with energy, innovation, and a culture of excellence, we welcome you to explore our career opportunities for both campus and experienced hires. Join us and be a part of creating the next milestone in the industry.** **Campus Recruitment**: https://hrrecommend.antgroup.com/guide.html?code=8uoP5mlus5DqQYbE_EnqcE2FD5JZH21MwvMUIb9mb6X3osXPuBraG54SyM8GLn_7 **Experienced Hires**: https://talent.antgroup.com/off-campus-position?positionId=1933830