The Byte team proposed the Lynx model: multi-modal LLMs understanding cognitive generation list SoTA

Current Large Language Models (LLMs) such as GPT4 exhibit excellent multi-modal capabilities in following open instructions given an image. However, the performance of these models heavily depends on the choices of network structure, training data, and training strategies, but these choices have not been widely discussed in the previous literature. In addition, there is currently a lack of suitable benchmarks to evaluate and compare these models, which limits the development of multimodal LLMs.

Picture

• Paper: https://arxiv.org/abs/2307.02469
• Website: https://lynx-llm.github.io/
• Code: https://github.com/bytedance/lynx-llm

In this article, the author conducts a systematic and comprehensive study on the training of such models from both quantitative and qualitative aspects. More than 20 variants were set up. For the network structure, different LLMs backbones and model designs were compared; for the training data, the impact of data and sampling strategies was studied; in terms of instructions, the effect of diverse prompts on the model's instruction following ability was explored. Influence. For benchmarks, the article first proposes an open visual question answering evaluation set Open-VQA including image and video tasks.

Based on the experimental conclusions, the author proposed Lynx, which shows the most accurate multi-modal understanding compared with the existing open source GPT4-style model capabilities while maintaining the best multi-modal generation capabilities.

Evaluation scheme

Unlike typical visual language tasks, the main challenge in evaluating GPT4-style models lies in balance Performance in two aspects: text generation ability and multimodal understanding accuracy . To solve this problem, the authors propose a new benchmark Open-VQA including video and image data, and conduct a comprehensive evaluation of current open source models.

Specifically, two quantitative evaluation schemes are adopted:

• Collect open visual question answering (Open-VQA) tests Set, which contains different categories of questions on objects, OCR, counting, reasoning, action recognition, time sequencing, etc. Unlike the VQA data set, which has standard answers, Open-VQA's answers are open-ended. To evaluate the performance on Open-VQA, GPT4 is used as the discriminator, and the results are 95% consistent with human evaluation.
• In addition, the author used the OwlEval data set provided by mPLUG-owl [1] to evaluate the text generation ability of the model. Although it only contains 50 pictures and 82 questions, it covers stories Generation, ad generation, code generation and other various problems, and recruit human annotators to score the performance of different models.

Conclusion

In order to deeply study the training strategy of multi-modal LLMs, the author mainly starts from the network structure (prefix fine-tuning/cross-attention force), training data (data selection and combination ratio), instructions (single instruction/diversified instructions), LLMs model (LLaMA [5]/Vicuna [6]), image pixels (420/224) and other aspects are set With more than twenty variations, the following main conclusions have been drawn through experiments:

• #Multimodal LLMs are less capable of following instructions than LLMs. For example, InstructBLIP [2] tends to generate short replies regardless of input instructions, while other models tend to generate long sentences regardless of instructions, which the authors believe is due to a lack of high-level responses. Resulting from quality and diverse multimodal instruction data.
• #The quality of training data is crucial to the performance of the model. Based on the results of experiments on different data, it was found that using a small amount of high-quality data performs better than using large-scale noisy data. The author believes that this is the difference between generative training and contrastive training, because generative training directly learns the conditional distribution of words rather than the similarity between text and images. Therefore, for better model performance, two things need to be met in terms of data: 1) contain high-quality smooth text; 2) text and image content are well aligned.
• Quests and prompts are critical to zero-shot capabilities. Using diverse tasks and instructions can improve the model's zero-shot generation ability on unknown tasks, which is consistent with observations in plain text models.
• It is important to balance correctness with language-generating ability. If the model is undertrained on downstream tasks (such as VQA), it is more likely to generate fabricated content that does not match the visual input; while if the model is overtrained on downstream tasks, it is more likely to generate fabricated content that does not match the visual input. Short answers will not be able to generate longer answers as directed by the user.
• Prefix-finetuning (PT) is currently the best solution for multi-modal adaptation of LLMs. In experiments, the model with prefix-finetuning structure can improve the ability to follow diverse instructions faster and is easier to train than the cross-attention (CA) model structure. (prefix-tuning and cross-attention are two model structures, see the Lynx model introduction section for details)

Lynx model

The author proposed Lynx(lynx) - a prefix-finetuning GPT4-style model with two-stage training. In the first stage, approximately 120M image-text pairs are used to align visual and language embeddings; in the second stage, 20 images or videos are used for multi-modal tasks and natural language processing (NLP) ) data to adjust the model's instruction-following capabilities.

Picture

The overall structure of the Lynx model is shown in Figure 1 above.

The visual input is processed by the visual encoder to obtain visual tokens (tokens) $$W_v$$. After mapping, it is spliced ​​with the instruction tokens $$W_l$$ as the input of LLMs. This structure is called "prefix-finetuning" in this article to distinguish it from the cross-attention structure used by Flamingo [3].

In addition, the author found that the training cost can be further reduced by adding Adapter (Adapter) after certain layers of frozen LLMs.

Model effect

The author evaluated the existing open source multi-modal LLMs model in Open-VQA, Mme [4] And the performance on OwlEval manual evaluation (results are shown in the chart below, and evaluation details are in the paper). It can be seen that the Lynx model has achieved the best performance in Open-VQA image and video understanding tasks, OwlEval manual evaluation and Mme Perception tasks. Among them, InstructBLIP also achieves high performance in most tasks, but its reply is too short. In comparison, in most cases, the Lynx model provides concise reasons to support the correct answer. Reply, which makes it more user-friendly (see the Cases display section below for some cases).

1. The indicator results on the Open-VQA image test set are shown in Table 1 below:

Picture

2. The indicator results on the Open-VQA video test set are shown in Table 2 below.

picture

3. Select the model with the top score in Open-VQA to conduct manual effect evaluation on the OwlEval evaluation set. The results are shown in Figure 4 above. It can be seen from the manual evaluation results that the Lynx model has the best language generation performance.

Picture

4. In the Mme benchmark test, the Perception class task achieved the best performance , among which 7 of 14 types of subtasks have the best performance. (See the appendix of the paper for detailed results)

Cases display

Open-VQA picture cases

OwlEval cases

Open-VQA video case

Summary

In this article, the author determined prefix-finetuning as the Open-VQA evaluation plan for the main structure of the Lynx model and open-ended answers. Experimental results show that the Lynx model performs the most accurate multi-modal understanding accuracy while maintaining the best multi-modal generation capabilities.

The above is the detailed content of The Byte team proposed the Lynx model: multi-modal LLMs understanding cognitive generation list SoTA. For more information, please follow other related articles on the PHP Chinese website!