Code LLMs have emerged as a specialized research field, with remarkable studies dedicated to enhancing model's coding capabilities through fine-tuning on pre-trained models. Previous fine-tuning approaches were typically tailored to specific downstream tasks or scenarios, which meant separate fine-tuning for each task, requiring extensive training resources and posing challenges in terms of deployment and maintenance. Furthermore, these approaches failed to leverage the inherent interconnectedness among different code-related tasks. To overcome these limitations, we present a multi-task fine-tuning framework, MFTCoder, that enables simultaneous and parallel fine-tuning on multiple tasks. By incorporating various loss functions, we effectively address common challenges in multi-task learning, such as data imbalance, varying difficulty levels, and inconsistent convergence speeds. Extensive experiments have conclusively demonstrated that our multi-task fine-tuning approach outperforms both individual fine-tuning on single tasks and fine-tuning on a mixed ensemble of tasks. Moreover, MFTCoder offers efficient training capabilities, including efficient data tokenization modes and parameter efficient fine-tuning (PEFT) techniques, resulting in significantly improved speed compared to traditional fine-tuning methods. MFTCoder seamlessly integrates with several mainstream open-source LLMs, such as CodeLLama and Qwen. Our MFTCoder fine-tuned CodeFuse-DeepSeek-33B claimed the top spot on the Big Code Models Leaderboard ranked by WinRate as of January 30, 2024. MFTCoder is open-sourced at https://github.com/codefuse-ai/MFTCOder
论文检索
输入标题、作者或关键词，从 642 篇学术成果中精准定位
Source localization aims to locate information diffusion sources only given the diffusion observation, which has attracted extensive attention in the past few years. Existing methods are mostly tailored for single networks and may not be generalized to handle more complex networks like cross-networks. Cross-network is defined as two interconnected networks, where one network's functionality depends on the other. Source localization on cross-networks entails locating diffusion sources on the source network by only giving the diffused observation in the target network. The task is challenging due to challenges including: 1) diffusion sources distribution modeling; 2) jointly considering both static and dynamic node features; and 3) heterogeneous diffusion patterns learning. In this work, we propose a novel method, namely CNSL, to handle the three primary challenges. Specifically, we propose to learn the distribution of diffusion sources through Bayesian inference and leverage disentangled encoders to learn static and dynamic node features separately. The learning objective is coupled with the cross-network information propagation estimation model to make the inference of diffusion sources considering the overall diffusion process. Additionally, we also provide two novel cross-network datasets collected by ourselves. Extensive experiments are conducted on both datasets to demonstrate the effectiveness of CNSL in handling the source localization on cross-networks.
Click-through rate (CTR) prediction is a crucial area of research in online advertising. While binary cross entropy (BCE) has been widely used as the optimization objective for treating CTR prediction as a binary classification problem, recent advancements have shown that combining BCE loss with an auxiliary ranking loss can significantly improve performance. However, the full effectiveness of this combination loss is not yet fully understood. In this paper, we uncover a new challenge associated with the BCE loss in scenarios where positive feedback is sparse: the issue of gradient vanishing for negative samples. We introduce a novel perspective on the effectiveness of the auxiliary ranking loss in CTR prediction: it generates larger gradients on negative samples, thereby mitigating the optimization difficulties when using the BCE loss only and resulting in improved classification ability. To validate our perspective, we conduct theoretical analysis and extensive empirical evaluations on public datasets. Additionally, we successfully integrate the ranking loss into Tencent's online advertising system, achieving notable lifts of 0.70% and 1.26% in Gross Merchandise Value (GMV) for two main scenarios. The code is openly accessible at: https://github.com/SkylerLinn/Understanding-the-Ranking-Loss.
Deep Bag-of-Words Model: An Efficient and Interpretable Relevance Architecture for Chinese E-Commerce
PDF ↗Text relevance or text matching of query and product is an essential technique for the e-commerce search system to ensure that the displayed products can match the intent of the query. Many studies focus on improving the performance of the relevance model in search system. Recently, pre-trained language models like BERT have achieved promising performance on the text relevance task. While these models perform well on the offline test dataset, there are still obstacles to deploy the pre-trained language model to the online system as their high latency. The two-tower model is extensively employed in industrial scenarios, owing to its ability to harmonize performance with computational efficiency. Regrettably, such models present an opaque ''black box'' nature, which prevents developers from making special optimizations. In this paper, we raise deep Bag-o f-Words (DeepBoW) model, an efficient and interpretable relevance architecture for Chinese e-commerce. Our approach proposes to encode the query and the product into the sparse BoW representation, which is a set of word-weight pairs. The weight means the important or the relevant score between the corresponding word and the raw text. The relevance score is measured by the accumulation of the matched word between the sparse BoW representation of the query and the product. Compared to popular dense distributed representation that usually suffers from the drawback of black-box, the most advantage of the proposed representation model is highly explainable and interventionable, which is a superior advantage to the deployment and operation of online search engines. Moreover, the online efficiency of the proposed model is even better than the most efficient inner product form of dense representation. The proposed model is experimented on three different datasets for learning the sparse BoW representations, including the human-annotation set, the search-log set and the click-through set. Then the models are evaluated by experienced human annotators. Both the auto metrics and the online evaluations show our DeepBoW model achieves competitive performance while the online inference is much more efficient than the other models. Our DeepBoW model has already deployed to the biggest Chinese e-commerce search engine Taobao and served the entire search traffic for over 6 months.
Text on historical maps contains valuable information providing georeferenced historical, political, and cultural contexts. However, text extraction from historical maps has been challenging due to the lack of (1) effective methods and (2) training data. Previous approaches use ad-hoc steps tailored to only specific map styles. Recent machine learning-based text spotters (e.g., for scene images) have the potential to solve these challenges because of their flexibility in supporting various types of text instances. However, these methods remain challenges in extracting precise image features for predicting every sub-component (boundary points and characters) in a text instance. This is critical because map text can be lengthy and highly rotated with complex backgrounds, posing difficulties in detecting relevant image features from a rough text region. This paper proposes PALETTE, an end-to-end text spotter for scanned historical maps of a wide variety. PALETTE introduces a novel hyper-local sampling module to explicitly learn localized image features around the target boundary points and characters of a text instance for detection and recognition. PALETTE also enables hyper-local positional embeddings to learn spatial interactions between boundary points and characters within and across text instances. In addition, this paper presents a novel approach to automatically generate synthetic map images, SYNTHMAP+, for training text spotters for historical maps. The experiment shows that PALETTE with SYNTHMAP+ outperforms SOTA text spotters on two new benchmark datasets of historical maps, particularly for long and angled text. We have deployed PALETTE with SYNTHMAP+ to process over 60,000 maps in the David Rumsey Historical Map collection and generated over 100 million text labels to support map searching.
Precise crop yield predictions are of national importance for ensuring food security and sustainable agricultural practices. While AI-for-science approaches have exhibited promising achievements in solving many scientific problems such as drug discovery, precipitation nowcasting, etc., the development of deep learning models for predicting crop yields is constantly hindered by the lack of an open and large-scale deep learning-ready dataset with multiple modalities to accommodate sufficient information. To remedy this, we introduce the CropNet dataset, the first terabyte-sized, publicly available, and multi-modal dataset specifically targeting climate change-aware crop yield predictions for the contiguous United States (U.S.) continent at the county level. Our CropNet dataset is composed of three modalities of data, i.e., Sentinel-2 Imagery, WRF-HRRR Computed Dataset, and USDA Crop Dataset, for over 2200 U.S. counties spanning 6 years (2017-2022), expected to facilitate researchers in developing versatile deep learning models for timely and precisely predicting crop yields at the county-level, by accounting for the effects of both short-term growing season weather variations and long-term climate change on crop yields. Besides, we develop the CropNet package, offering three types of APIs, for facilitating researchers in downloading the CropNet data on the fly over the time and region of interest, and flexibly building their deep learning models for accurate crop yield predictions. Extensive experiments have been conducted on our CropNet dataset via employing various types of deep learning solutions, with the results validating the general applicability and the efficacy of the CropNet dataset in climate change-aware crop yield predictions. We have officially released our CropNet dataset on Hugging Face Datasets https://huggingface.co/datasets/CropNet/CropNet and our CropNet package on the Python Package Index (PyPI) https://pypi.org/project/cropnet. Code and tutorials are available at https://github.com/fudong03/CropNet.
Harvesting Efficient On-Demand Order Pooling from Skilled Couriers: Enhancing Graph Representation Learning for Refining Real-time Many-to-One Assignments
PDF ↗The recent past has witnessed a notable surge in on-demand food delivery (OFD) services, offering delivery fulfillment within dozens of minutes after an order is placed. In OFD, pooling multiple orders for simultaneous delivery in real-time order assignment is a pivotal efficiency source, which may in turn extend delivery time. Constructing high-quality order pooling to harmonize platform efficiency with the experiences of consumers and couriers, is crucial to OFD platforms. However, the complexity and real-time nature of order assignment, making extensive calculations impractical, significantly limit the potential for order consolidation. Moreover, offline environment is frequently riddled with unknown factors, posing challenges for the platform's perceptibility and pooling decisions. Nevertheless, delivery behaviors of skilled couriers (SCs) who know the environment well, can improve system awareness and effectively inform decisions. Hence a SC delivery network (SCDN) is constructed, based on an enhanced attributed heterogeneous network embedding approach tailored for OFD. It aims to extract features from rich temporal and spatial information, and uncover the latent potential for order combinations embedded within SC trajectories. Accordingly, the vast search space of order assignment can be effectively pruned through scalable similarity calculations of low-dimensional vectors, making comprehensive and high-quality pooling outcomes more easily identified in real time. In addition, the acquired embedding outcomes highlight promising subspaces embedded within this space, i.e., scale-effect hotspot areas, which can offer significant potential for elevating courier efficiency. SCDN has now been deployed in Meituan dispatch system. Online tests reveal that with SCDN, the pooling quality and extent have been greatly improved. And our system can boost couriers' efficiency by 45-55% during noon peak hours, while upholding the timely delivery commitment.
Spatio-temporal prediction aims to forecast and gain insights into the ever-changing dynamics of urban environments across both time and space. Its purpose is to anticipate future patterns, trends, and events in diverse facets of urban life, including transportation, population movement, and crime rates. Although numerous efforts have been dedicated to developing neural network techniques for accurate predictions on spatio-temporal data, it is important to note that many of these methods heavily depend on having sufficient labeled data to generate precise spatio-temporal representations. Unfortunately, the issue of data scarcity is pervasive in practical urban sensing scenarios. In certain cases, it becomes challenging to collect any labeled data from downstream scenarios, intensifying the problem further. Consequently, it becomes necessary to build a spatio-temporal model that can exhibit strong generalization capabilities across diverse spatio-temporal learning scenarios. Taking inspiration from the remarkable achievements of large language models (LLMs), our objective is to create a spatio-temporal LLM that can exhibit exceptional generalization capabilities across a wide range of downstream urban tasks. To achieve this objective, we present the UrbanGPT, which seamlessly integrates a spatio-temporal dependency encoder with the instruction-tuning paradigm. This integration enables LLMs to comprehend the complex inter-dependencies across time and space, facilitating more comprehensive and accurate predictions under data scarcity. To validate the effectiveness of our approach, we conduct extensive experiments on various public datasets, covering different spatio-temporal prediction tasks. The results consistently demonstrate that our UrbanGPT, with its carefully designed architecture, consistently outperforms state-of-the-art baselines. These findings highlight the potential of building large language models for spatio-temporal learning, particularly in zero-shot scenarios where labeled data is scarce. The code and data are available at: https://github.com/HKUDS/UrbanGPT.
Product search is an important service on Taobao, the largest e-commerce platform in China. Through this service, users can easily find products relevant to their specific needs. Coping with billion-size query loads, Taobao product search has traditionally relied on classical term-based retrieval models due to their powerful and interpretable indexes. In essence, efficient retrieval hinges on the proper storage of the inverted index. Recent successes involve reducing the size (pruning) of the inverted index but the construction and deployment of lossless static index pruning in practical product search still pose non-trivial challenges. In this work, we introduce a novel SM art INDexing (SMIND) solution in Taobao product search. SMIND is designed to reduce information loss during the static pruning process by incorporating user search preferences. Specifically, we first construct "user-query-item'' hypergraphs for four different search preferences, namely purchase, click, exposure, and relevance. Then, we develop an efficient TermRank algorithm applied to these hypergraphs, to preserve relevant items based on specific user preferences during the pruning of the inverted indexer. Our approach offers fresh insights into the field of product search, emphasizing that term dependencies in user search preferences go beyond mere text relevance. Moreover, to address the vocabulary mismatch problem inherent in term-based models, we also incorporate an multi-granularity semantic retrieval model to facilitate semantic matching. Empirical results from both offline evaluation and online A/B tests showcase the superiority of SMIND over state-of-the-art methods, especially in commerce metrics with significant improvements of 1.34% in Pay Order Count and 1.50% in Gross Merchandise Value. Besides, SMIND effectively mitigates the Matthew effect of user queries and has been in service for hundreds of millions of daily users since November 2022.
Graph-based fraud detection has widespread application in modern industry scenarios, such as spam review and malicious account detection. While considerable efforts have been devoted to designing adequate fraud detectors, the interpretability of their results has often been overlooked. Previous works have attempted to generate explanations for specific instances using post-hoc explaining methods such as a GNNExplainer. However, post-hoc explanations can not facilitate the model predictions and the computational cost of these methods cannot meet practical requirements, thus limiting their application in real-world scenarios. To address these issues, we propose SEFraud, a novel graph-based self-explainable fraud detection framework that simultaneously tackles fraud detection and result in interpretability. Concretely, SEFraud first leverages customized heterogeneous graph transformer networks with learnable feature masks and edge masks to learn expressive representations from the informative heterogeneously typed transactions. A new triplet loss is further designed to enhance the performance of mask learning. Empirical results on various datasets demonstrate the effectiveness of SEFraud as it shows considerable advantages in both the fraud detection performance and interpretability of prediction results. Specifically, SEFraud achieves the most significant improvement with 8.6% on AUC and 8.5% on Recall over the second best on fraud detection, as well as an average of 10x speed-up regarding the inference time. Last but not least, SEFraud has been deployed and offers explainable fraud detection service for the largest bank in China, Industrial and Commercial Bank of China Limited (ICBC). Results collected from the production environment of ICBC show that SEFraud can provide accurate detection results and comprehensive explanations that align with the expert business understanding, confirming its efficiency and applicability in large-scale online services.
Pathogenic chromosome abnormalities are very common among the general population. While numerical chromosome abnormalities can be quickly and precisely detected, structural chromosome abnormalities are far more complex and typically require considerable efforts by human experts for identification. This paper focuses on investigating the modeling of chromosome features and the identification of chromosomes with structural abnormalities. Most existing data-driven methods concentrate on a single chromosome and consider each chromosome independently, overlooking the crucial aspect of homologous chromosomes. In normal cases, homologous chromosomes share identical structures, with the exception that one of them is abnormal. Therefore, we propose an adaptive method to align homologous chromosomes and diagnose structural abnormalities through homologous similarity. Inspired by the process of human expert diagnosis, we incorporate information from multiple pairs of homologous chromosomes simultaneously, aiming to reduce noise disturbance and improve prediction performance. Extensive experiments on real-world datasets validate the effectiveness of our model compared to baselines.
The diversity of recommendation is equally crucial as accuracy in improving user experience. Existing studies, e.g., Determinantal Point Process (DPP) and Maximal Marginal Relevance (MMR), employ a greedy paradigm to iteratively select items that optimize both accuracy and diversity. However, prior methods typically exhibit quadratic complexity, limiting their applications to the re-ranking stage and are not applicable to other recommendation stages with a larger pool of candidate items, such as the pre-ranking and ranking stages. In this paper, we propose Contextual Distillation Model (CDM), an efficient recommendation model that addresses diversification, suitable for the deployment in all stages of industrial recommendation pipelines. Specifically, CDM utilizes the candidate items in the same user request as context to enhance the diversification of the results. We propose a contrastive context encoder that employs attention mechanisms to model both positive and negative contexts. For the training of CDM, we compare each target item with its context embedding and utilize the knowledge distillation framework to learn the win probability of each target item under the MMR algorithm, where the teacher is derived from MMR outputs. During inference, ranking is performed through a linear combination of the recommendation and student model scores, ensuring both diversity and efficiency. We perform offline evaluations on two industrial datasets and conduct online A/B test of CDM on the short-video platform KuaiShou. The considerable enhancements observed in both recommendation quality and diversity, as shown by metrics, provide strong superiority for the effectiveness of CDM.
Large language models (LLMs) have fueled many intelligent web agents, but most existing ones perform far from satisfying in real-world web navigation tasks due to three factors: (1) the complexity of HTML text data (2) versatility of actions on webpages, and (3) task difficulty due to the open-domain nature of the web. In light of these challenges, we develop the open AutoWebGLM based on ChatGLM3-6B. AutoWebGLM can serve as a powerful automated web navigation agent that outperform GPT-4. Inspired by human browsing patterns, we first design an HTML simplification algorithm to represent webpages with vital information preserved succinctly. We then employ a hybrid human-AI method to build web browsing data for curriculum training. Finally, we bootstrap the model by reinforcement learning and rejection sampling to further facilitate webpage comprehension, browser operations, and efficient task decomposition by itself. For comprehensive evaluation, we establish a bilingual benchmark---AutoWebBench---for real-world web navigation tasks. We evaluate AutoWebGLM across diverse web navigation benchmarks, demonstrating its potential to tackle challenging tasks in real environments. Related code, model, and data are released at https://github.com/THUDM/AutoWebGLM.
The Sustainable Development Goals (SDGs) of the United Nations provide a blueprint of a better future by "leaving no one behind", and, to achieve the SDGs by 2030, poor countries require immense volumes of development aid. In this paper, we develop a causal machine learning framework for predicting heterogeneous treatment effects of aid disbursements to inform effective aid allocation. Specifically, our framework comprises three components: (i) a balancing autoencoder that uses representation learning to embed high-dimensional country characteristics while addressing treatment selection bias; (ii) a counterfactual generator to compute counterfactual outcomes for varying aid volumes to address small sample-size settings; and (iii) an inference model that is used to predict heterogeneous treatment-response curves. We demonstrate the effectiveness of our framework using data with official development aid earmarked to end HIV/AIDS in 105 countries, amounting to more than USD 5.2 billion. For this, we first show that our framework successfully computes heterogeneous treatment-response curves using semi-synthetic data. Then, we demonstrate our framework using real-world HIV data. Our framework points to large opportunities for a more effective aid allocation, suggesting that the total number of new HIV infections could be reduced by up to 3.3% (~50,000 cases) compared to the current allocation practice.
When a customer overdraws their bank account and their balance is negative they are assessed an overdraft fee. Americans pay approximately $15 billion in unnecessary overdraft fees a year, often in $35 increments; users of the Mint personal finance app pay approximately $250 million in fees a year in particular. These overdraft fees are an excessive financial burden and lead to cascading overdraft fees trapping customers in financial hardship. To address this problem, we have created an ML-driven overdraft early warning system (ODEWS) that assesses a customer's risk of overdrafting within the next week using their banking and transaction data in the Mint app. At-risk customers are sent an alert so they can take steps to avoid the fee, ultimately changing their behavior and financial habits. The system deployed resulted in a $3 million savings in overdraft fees for Mint customers compared to a control group. Moreover, the methodology outlined here is part of a greater effort to provide ML-driven personalized financial advice to help our members know, grow, and protect their net worth, ultimately, achieving their financial goals.
FederatedScope-LLM: A Comprehensive Package for Fine-tuning Large Language Models in Federated Learning
PDF ↗Large language models (LLMs) have demonstrated great capabilities in various natural language understanding and generation tasks. These pre-trained LLMs can be further improved for specific downstream tasks by fine-tuning. However, the adoption of LLM in real-world applications can be hindered by privacy concerns and the resource-intensive nature of model training and fine-tuning. When multiple entities have similar interested tasks but cannot directly share their local data due to privacy regulations, federated learning (FL) is a mainstream solution to leverage the data of different entities. Besides avoiding direct data sharing, FL can also achieve rigorous data privacy protection, model intelligent property protection, and model customization via composition with different techniques. Despite the aforementioned advantages of FL, fine-tuning LLMs in FL settings still lacks adequate support from the existing frameworks and, therefore, faces challenges in optimizing the consumption of significant communication and computational resources, preparing various data for different tasks, and satisfying diverse information protection demands. In this paper, we discuss these challenges and introduce our package FederatedScope-LLM (FS-LLM) as a main contribution, which consists: (1) We build a complete end-to-end benchmarking pipeline under real-world scenarios, automizing the processes of dataset preprocessing, federated fine-tuning execution or simulation, and performance evaluation; (2) We provide comprehensive and off-the-shelf federated parameter-efficient fine-tuning (PEFT) algorithm implementations and versatile programming interfaces for future extension, enhancing the capabilities of LLMs in FL scenarios with low communication and computation costs, even without accessing the full model; (3) We adopt several accelerating and resource-efficient operators, and provide flexible pluggable sub-routines for interdisciplinary study. We conduct extensive and reproducible experiments to show the effectiveness of FS-LLM and benchmark advanced LLMs with PEFT algorithms in FL. We release FS-LLM at https://github.com/alibaba/FederatedScope/tree/llm.
The online advertising market, with its thousands of auctions run per second, presents a daunting challenge for advertisers who wish to optimize their spend under a budget constraint. Thus, advertising platforms typically provide automated agents to their customers, which act on their behalf to bid for impression opportunities in real time at scale. Because these proxy agents are owned by the platform but use advertiser funds to operate, there is a strong practical need to balance reliability and explainability of the agent with optimizing power. We propose a generalizable approach to optimizing bidding policies in production environments by learning from real data using offline reinforcement learning. This approach can be used to optimize any differentiable base policy (practically, a heuristic policy based on principles which the advertiser can easily understand), and only requires data generated by the base policy itself. We use a hybrid agent architecture that combines arbitrary base policies with deep neural networks, where only the optimized base policy parameters are eventually deployed, and the neural network part is discarded after training. We demonstrate that such an architecture achieves statistically significant performance gains in both simulated and at-scale production bidding environments compared to the default production bidding policy. Our approach does not incur additional infrastructure, safety, or explainability costs, as it directly optimizes the parameters of existing production routines without necessarily replacing them with black box-style models like neural networks.
A/B tests, or online controlled experiments, are used heavily in the software industry to evaluate implementations of ideas, as the paradigm is the gold standard in science for establishing causality: the changes introduced in the treatment caused the changes to the metrics of interest with high probability. What distinguishes software experiments, or A/B tests, from experiments in many other domains is the scale (e.g., over 100 experiment treatments may launch on a given workday in large companies) and the effect sizes that matter to the business are small (e.g., a 3% improvement to conversion rate from a single experiment is a cause for celebration). The humbling reality is that most experiments fail to improve key metrics, and success rates of only about 10-20% are most common. With low success rates, the industry standard alpha threshold of 0.05 implies a high probability of false positives. We begin with motivation about why false positives are expensive in many software domains. We then offer several approaches to estimate the true success rate of experiments, given the observed "win" rate (statistically significant positive improvements), and show examples from Expedia and Optimizely. We offer a modified procedure for experimentation, based in sequential group testing, that selectively extends experiments to reduce false positives, increase power, at a small increase to runtime. We conclude with a discussion of the difference between ideas and experiments in practice, terms that are often incorrectly used interchangeably.
Hierarchical time-series forecasting (HTSF) is an important problem for many real-world business applications where the goal is to simultaneously forecast multiple time-series that are related to each other via a hierarchical relation. Recent works, however, do not address two important challenges that are typically observed in many demand forecasting applications at large companies. First, many time-series at lower levels of the hierarchy have high sparsity i.e., they have a significant number of zeros. Most HTSF methods do not address this varying sparsity across the hierarchy. Further, they do not scale well to the large size of the real-world hierarchy typically unseen in benchmarks used in literature. We resolve both these challenges by proposing HAILS, a novel probabilistic hierarchical model that enables accurate and calibrated probabilistic forecasts across the hierarchy by adaptively modeling sparse and dense time-series with different distributional assumptions and reconciling them to adhere to hierarchical constraints. We show the scalability and effectiveness of our methods by evaluating them against real-world demand forecasting datasets. We deploy HAILS at a large chemical manufacturing company for a product demand forecasting application with over ten thousand products and observe a significant 8.5% improvement in forecast accuracy and 23% better improvement for sparse time-series. The enhanced accuracy and scalability make HAILS a valuable tool for improved business planning and customer experience.
RJUA-MedDQA: A Multimodal Benchmark for Medical Document Question Answering and Clinical Reasoning
PDF ↗Recent advancements in Large Language Models (LLMs) and Large Multi-modal Models (LMMs) have shown potential in various medical applications, such as Intelligent Medical Diagnosis. Although impressive results have been achieved, we find that existing benchmarks do not reflect the complexity of real medical reports and specialized in-depth reasoning capabilities. In this work, we establish a comprehensive benchmark in the field of medical specialization and introduced RJUA-MedDQA, which contains 2000 real-world Chinese medical report images poses several challenges: comprehensively interpreting imgage content across a wide variety of challenging layouts, possessing the numerical reasoning ability to identify abnormal indicators and demonstrating robust clinical reasoning ability to provide the statement of disease diagnosis, status and advice based on a collection of medical contexts. We carefully design the data generation pipeline and proposed the Efficient Structural Restoration Annotation (ESRA) Method, aimed at restoring textual and tabular content in medical report images. This method substantially enhances annotation efficiency, doubling the productivity of each annotator, and yields a 26.8% improvement in accuracy. We conduct extensive evaluations, including few-shot assessments of 5 LMMs which are capable of solving Chinese medical QA tasks. To further investigate the limitations and potential of current LMMs, we conduct comparative experiments on a set of strong LLMs by using image-text generated by ESRA method. We report the performance of baselines and offer several observations: (1) The overall performance of existing LMMs is still limited; however LMMs more robust to low-quality and diverse-structured images compared to LLMs. (3) Reasoning across context and image content present significant challenges. We hope this benchmark helps the community make progress on these challenging tasks in multi-modal medical document understanding and facilitate its application in healthcare. Our dataset will be publicly available for noncommercial use at https://github.com/Alipay-Med/medDQA_benchmark.git