Liang Zhao

Tianwen Wei, Liang Zhao, Lichang Zhang et al.

In this technical report, we present Skywork-13B, a family of large language models (LLMs) trained on a corpus of over 3.2 trillion tokens drawn from both English and Chinese texts. This bilingual foundation model is the most extensively trained and openly published LLMs of comparable size to date. We introduce a two-stage training methodology using a segmented corpus, targeting general purpose training and then domain-specific enhancement training, respectively. We show that our model not only excels on popular benchmarks, but also achieves \emph{state of the art} performance in Chinese language modeling on diverse domains. Furthermore, we propose a novel leakage detection method, demonstrating that test data contamination is a pressing issue warranting further investigation by the LLM community. To spur future research, we release Skywork-13B along with checkpoints obtained during intermediate stages of the training process. We are also releasing part of our SkyPile corpus, a collection of over 150 billion tokens of web text, which is the largest high quality open Chinese pre-training corpus to date. We hope Skywork-13B and our open corpus will serve as a valuable open-source resource to democratize access to high-quality LLMs.

Ailin Huang, Boyong Wu, Bruce Wang et al.

Real-time speech interaction, serving as a fundamental interface for human-machine collaboration, holds immense potential. However, current open-source models face limitations such as high costs in voice data collection, weakness in dynamic control, and limited intelligence. To address these challenges, this paper introduces Step-Audio, the first production-ready open-source solution. Key contributions include: 1) a 130B-parameter unified speech-text multi-modal model that achieves unified understanding and generation, with the Step-Audio-Chat version open-sourced; 2) a generative speech data engine that establishes an affordable voice cloning framework and produces the open-sourced lightweight Step-Audio-TTS-3B model through distillation; 3) an instruction-driven fine control system enabling dynamic adjustments across dialects, emotions, singing, and RAP; 4) an enhanced cognitive architecture augmented with tool calling and role-playing abilities to manage complex tasks effectively. Based on our new StepEval-Audio-360 evaluation benchmark, Step-Audio achieves state-of-the-art performance in human evaluations, especially in terms of instruction following. On open-source benchmarks like LLaMA Question, shows 9.3% average performance improvement, demonstrating our commitment to advancing the development of open-source multi-modal language technologies. Our code and models are available at https://github.com/stepfun-ai/Step-Audio.

Haoran Wei, Lingyu Kong, Jinyue Chen et al.

Playing Large Vision Language Models (LVLMs) in 2023 is trendy among the AI community. However, the relatively large number of parameters (more than 7B) of popular LVLMs makes it difficult to train and deploy on consumer GPUs, discouraging many researchers with limited resources. Imagine how cool it would be to experience all the features of current LVLMs on an old GTX1080ti (our only game card). Accordingly, we present Vary-toy in this report, a small-size Vary along with Qwen-1.8B as the base ``large'' language model. In Vary-toy, we introduce an improved vision vocabulary, allowing the model to not only possess all features of Vary but also gather more generality. Specifically, we replace negative samples of natural images with positive sample data driven by object detection in the procedure of generating vision vocabulary, more sufficiently utilizing the capacity of the vocabulary network and enabling it to efficiently encode visual information corresponding to natural objects. For experiments, Vary-toy can achieve 65.6% ANLS on DocVQA, 59.1% accuracy on ChartQA, 88.1% accuracy on RefCOCO, and 29% on MMVet. The code will be publicly available on the homepage.

Mingchen Li, Chen Ling, Rui Zhang et al.

Zero-shot link prediction (ZSLP) on knowledge graphs aims at automatically identifying relations between given entities. Existing methods primarily employ auxiliary information to predict tail entity given head entity and its relation, yet face challenges due to the occasional unavailability of such detailed information and the inherent simplicity of predicting tail entities based on semantic similarities. Even though Large Language Models (LLMs) offer a promising solution to predict unobserved relations between the head and tail entity in a zero-shot manner, their performance is still restricted due to the inability to leverage all the (exponentially many) paths' information between two entities, which are critical in collectively indicating their relation types. To address this, in this work, we introduce a Condensed Transition Graph Framework for Zero-Shot Link Prediction (CTLP), which encodes all the paths' information in linear time complexity to predict unseen relations between entities, attaining both efficiency and information preservation. Specifically, we design a condensed transition graph encoder with theoretical guarantees on its coverage, expressiveness, and efficiency. It is learned by a transition graph contrastive learning strategy. Subsequently, we design a soft instruction tuning to learn and map the all-path embedding to the input of LLMs. Experimental results show that our proposed CTLP method achieves state-of-the-art performance on three standard ZSLP datasets

Liang Zhao, Tianwen Wei, Liang Zeng et al.

We introduce LongSkywork, a long-context Large Language Model (LLM) capable of processing up to 200,000 tokens. We provide a training recipe for efficiently extending context length of LLMs. We identify that the critical element in enhancing long-context processing capability is to incorporate a long-context SFT stage following the standard SFT stage. A mere 200 iterations can convert the standard SFT model into a long-context model. To reduce the effort in collecting and annotating data for long-context language modeling, we develop two novel methods for creating synthetic data. These methods are applied during the continual pretraining phase as well as the Supervised Fine-Tuning (SFT) phase, greatly enhancing the training efficiency of our long-context LLMs. Our findings suggest that synthetic long-context SFT data can surpass the performance of data curated by humans to some extent. LongSkywork achieves outstanding performance on a variety of long-context benchmarks. In the Needle test, a benchmark for long-context information retrieval, our models achieved perfect accuracy across multiple context spans. Moreover, in realistic application scenarios, LongSkywork-13B demonstrates performance on par with Claude2.1, the leading long-context model, underscoring the effectiveness of our proposed methods.

Yuyang Gao, Giorgio A. Ascoli, Liang Zhao

Continual learning is considered a promising step towards next-generation Artificial Intelligence (AI), where deep neural networks (DNNs) make decisions by continuously learning a sequence of different tasks akin to human learning processes. It is still quite primitive, with existing works focusing primarily on avoiding (catastrophic) forgetting. However, since forgetting is inevitable given bounded memory and unbounded task loads, 'how to reasonably forget' is a problem continual learning must address in order to reduce the performance gap between AIs and humans, in terms of 1) memory efficiency, 2) generalizability, and 3) robustness when dealing with noisy data. To address this, we propose a novel ScheMAtic memory peRsistence and Transience (SMART) framework for continual learning with external memory that builds on recent advances in neuroscience. The efficiency and generalizability are enhanced by a novel long-term forgetting mechanism and schematic memory, using sparsity and 'backward positive transfer' constraints with theoretical guarantees on the error bound. Robust enhancement is achieved using a novel short-term forgetting mechanism inspired by background information-gated learning. Finally, an extensive experimental analysis on both benchmark and real-world datasets demonstrates the effectiveness and efficiency of our model.

Boxiang Dong, Aparna S. Varde, Danilo Stevanovic et al.

Handwriting recognition is of crucial importance to both Human Computer Interaction (HCI) and paperwork digitization. In the general field of Optical Character Recognition (OCR), handwritten Chinese character recognition faces tremendous challenges due to the enormously large character sets and the amazing diversity of writing styles. Learning an appropriate distance metric to measure the difference between data inputs is the foundation of accurate handwritten character recognition. Existing distance metric learning approaches either produce unacceptable error rates, or provide little interpretability in the results. In this paper, we propose an interpretable distance metric learning approach for handwritten Chinese character recognition. The learned metric is a linear combination of intelligible base metrics, and thus provides meaningful insights to ordinary users. Our experimental results on a benchmark dataset demonstrate the superior efficiency, accuracy and interpretability of our proposed approach.

Liang Zhao, Hexin Cao, Yunsong Zhao

A new method for Text-to-SQL parsing, Grammar Pre-training (GP), is proposed to decode deep relations between question and database. Firstly, to better utilize the information of databases, a random value is added behind a question word which is recognized as a column, and the new sentence serves as the model input. Secondly, initialization of vectors for decoder part is optimized, with reference to the former encoding so that question information can be concerned. Finally, a new approach called flooding level is adopted to get the non-zero training loss which can generalize better results. By encoding the sentence with GRAPPA and RAT-SQL model, we achieve better performance on spider, a cross-DB Text-to-SQL dataset (72.8 dev, 69.8 test). Experiments show that our method is easier to converge during training and has excellent robustness.

Liang Zhao, Jingjing Xu, Junyang Lin et al.

Long text generation is an important but challenging task.The main problem lies in learning sentence-level semantic dependencies which traditional generative models often suffer from. To address this problem, we propose a Multi-hop Reasoning Generation (MRG) approach that incorporates multi-hop reasoning over a knowledge graph to learn semantic dependencies among sentences. MRG consists of twoparts, a graph-based multi-hop reasoning module and a path-aware sentence realization module. The reasoning module is responsible for searching skeleton paths from a knowledge graph to imitate the imagination process in the human writing for semantic transfer. Based on the inferred paths, the sentence realization module then generates a complete sentence. Unlike previous black-box models, MRG explicitly infers the skeleton path, which provides explanatory views tounderstand how the proposed model works. We conduct experiments on three representative tasks, including story generation, review generation, and product description generation. Automatic and manual evaluation show that our proposed method can generate more informative and coherentlong text than strong baselines, such as pre-trained models(e.g. GPT-2) and knowledge-enhanced models.

Liang Zhao, Yang Wang, Daxiang Dong et al.

In this paper we propose to solve an important problem in recommendation -- user cold start, based on meta leaning method. Previous meta learning approaches finetune all parameters for each new user, which is both computing and storage expensive. In contrast, we divide model parameters into fixed and adaptive parts and develop a two-stage meta learning algorithm to learn them separately. The fixed part, capturing user invariant features, is shared by all users and is learned during offline meta learning stage. The adaptive part, capturing user specific features, is learned during online meta learning stage. By decoupling user invariant parameters from user dependent parameters, the proposed approach is more efficient and storage cheaper than previous methods. It also has potential to deal with catastrophic forgetting while continually adapting for streaming coming users. Experiments on production data demonstrates that the proposed method converges faster and to a better performance than baseline methods. Meta-training without online meta model finetuning increases the AUC from 72.24% to 74.72% (2.48% absolute improvement). Online meta training achieves a further gain of 2.46\% absolute improvement comparing with offline meta training.

Yuanpeng Li, Liang Zhao, Jianyu Wang et al.

Compositional generalization is a basic mechanism in human language learning, but current neural networks lack such ability. In this paper, we conduct fundamental research for encoding compositionality in neural networks. Conventional methods use a single representation for the input sentence, making it hard to apply prior knowledge of compositionality. In contrast, our approach leverages such knowledge with two representations, one generating attention maps, and the other mapping attended input words to output symbols. We reduce the entropy in each representation to improve generalization. Our experiments demonstrate significant improvements over the conventional methods in five NLP tasks including instruction learning and machine translation. In the SCAN domain, it boosts accuracies from 14.0% to 98.8% in Jump task, and from 92.0% to 99.7% in TurnLeft task. It also beats human performance on a few-shot learning task. We hope the proposed approach can help ease future research towards human-level compositional language learning.

Liang Zhao, Wei Xu

In this paper we present our scientific discovery that good representation can be learned via continuous attention during the interaction between Unsupervised Learning(UL) and Reinforcement Learning(RL) modules driven by intrinsic motivation. Specifically, we designed intrinsic rewards generated from UL modules for driving the RL agent to focus on objects for a period of time and to learn good representations of objects for later object recognition task. We evaluate our proposed algorithm in both with and without extrinsic reward settings. Experiments with end-to-end training in simulated environments with applications to few-shot object recognition demonstrated the effectiveness of the proposed algorithm.

He Bai, Yu Zhou, Jiajun Zhang et al.

To deploy a spoken language understanding (SLU) model to a new language, language transferring is desired to avoid the trouble of acquiring and labeling a new big SLU corpus. Translating the original SLU corpus into the target language is an attractive strategy. However, SLU corpora consist of plenty of semantic labels (slots), which general-purpose translators cannot handle well, not to mention additional culture differences. This paper focuses on the language transferring task given a tiny in-domain parallel SLU corpus. The in-domain parallel corpus can be used as the first adaptation on the general translator. But more importantly, we show how to use reinforcement learning (RL) to further finetune the adapted translator, where translated sentences with more proper slot tags receive higher rewards. We evaluate our approach on Chinese to English language transferring for SLU systems. The experimental results show that the generated English SLU corpus via adaptation and reinforcement learning gives us over 97% in the slot F1 score and over 84% accuracy in domain classification. It demonstrates the effectiveness of the proposed language transferring method. Compared with naive translation, our proposed method improves domain classification accuracy by relatively 22%, and the slot filling F1 score by relatively more than 71%.

Yang Wang, Yi Yang, Zhenheng Yang et al.

It has been recently shown that a convolutional neural network can learn optical flow estimation with unsupervised learning. However, the performance of the unsupervised methods still has a relatively large gap compared to its supervised counterpart. Occlusion and large motion are some of the major factors that limit the current unsupervised learning of optical flow methods. In this work we introduce a new method which models occlusion explicitly and a new warping way that facilitates the learning of large motion. Our method shows promising results on Flying Chairs, MPI-Sintel and KITTI benchmark datasets. Especially on KITTI dataset where abundant unlabeled samples exist, our unsupervised method outperforms its counterpart trained with supervised learning.

Zhenheng Yang, Peng Wang, Wei Xu et al.

Learning to reconstruct depths in a single image by watching unlabeled videos via deep convolutional network (DCN) is attracting significant attention in recent years. In this paper, we introduce a surface normal representation for unsupervised depth estimation framework. Our estimated depths are constrained to be compatible with predicted normals, yielding more robust geometry results. Specifically, we formulate an edge-aware depth-normal consistency term, and solve it by constructing a depth-to-normal layer and a normal-to-depth layer inside of the DCN. The depth-to-normal layer takes estimated depths as input, and computes normal directions using cross production based on neighboring pixels. Then given the estimated normals, the normal-to-depth layer outputs a regularized depth map through local planar smoothness. Both layers are computed with awareness of edges inside the image to help address the issue of depth/normal discontinuity and preserve sharp edges. Finally, to train the network, we apply the photometric error and gradient smoothness for both depth and normal predictions. We conducted experiments on both outdoor (KITTI) and indoor (NYUv2) datasets, and show that our algorithm vastly outperforms state of the art, which demonstrates the benefits from our approach.