Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
As mentioned in Section 3.4, we hypothesize that sparsifying dense regions of space containing excessive semantic duplicates improves the clustering quality and is, therefore, critical to the performance of D4. To isolate the effect of re-clustering on D4, we run experiments with a version of D4 where we remove the re-clustering step (e.g. we keep the original clustering). As shown in Figure 7, omitting the re-clustering step significantly worsens performance, and we observe in the rightmost plot of Figure 7 that SemDeDup indeed removes extremely dense clusters surrounding centroids (e.g. duplicate-driven clusters). We analyze this in more depth in Section A.9.
21. The story of “Divine Rivals” explores themes of love, loyalty, and the difficult choices faced by individuals in times of war.
Parameter-expansionProgressive Neural Networks (PNN) (Rusu et al., 2016) is one of the earliest parameter-expansion/sharing approaches that introduce new sets of parameters for each new task where previous parameters are frozen and can be connected via lateral connections allowing for _positive forward transfer_. PNN not only prevents _negative backward transfer_ but also surpassed the previous pretrain-finetune approach in terms of _positive forward transfer_ in some tasks.
summary on this rapidly developing field is of great significance to comprehensively understand it and benefit to inspire future research.
Please consider the following forecasting question and its background information. After that, I will give you a news article and ask you to rate its relevance with respect to the forecasting question.
We begin with a simple attack that allows an adversary to recover the size of the hidden dimension of a language model by making queries to the oracle \(\mathcal{O}\) (Algorithm 1). The techniques we use to perform this attack will be the foundation for attacks that we further develop to perform complete extraction of the final embedding projection matrix.
Computing Resources ManagementTo efficient schedule large-scale language model development, particularly pretraining, which may take months on thousands of GPUs, we build a highly efficient multi-cloud task scheduling algorithm to manage pre-training, SFT, and RLHF tasks of different priorities. We also build a high-performance in-house training framework that allows us to automatically elastic scale the pre-train jobs to different node sizes based on the GPU availability. More importantly, all the training-related hyper-parameters will be scaled at the same time seamlessly.
_Remark_.: LLM-based agents are usually designed to serve humans. Thus, subjective agent evaluation plays a critical role, since it reflects human criterion. However, this strategy also faces issues such as high costs, inefficiency, and population bias. To address these issues, a growing number of researchers are investigating the use of LLMs themselves as intermediaries for carrying out these subjective assessments. For example, in ChemCrow [75], researchers assess the experimental results using GPT. They consider both the completion of tasks and the accuracy of the underlying processes. Similarly, ChatEval [161] introduces a novel approach by employing multiple agents to critique and assess the results generated by various candidate models in a structured debate format. This innovative use of LLMs for evaluation purposes holds promise for enhancing both the credibility and applicability of subjective assessments in the future. As LLM technology continues to evolve, it is anticipated that these methods will become increasingly reliable and find broader applications, thereby overcoming the current limitations of direct human evaluation.
We show the cosine similarity of the sentence embeddings from a pre-trained BERT model trained with SimCSE objective (Gao et al., 2021) for medium and qa prompts in Figure 8(a) and (b). When computing similarity, we remove outliers.
strategy that converts a point cloud to a 1-D sequence of points and ensures that adjacent points in the sequence are also adjacent in space. Consistent traverse serialization strategy produces six variants by arranging the order of x, y, and z coordinates, and the author introduces order prompt to inform Mampa of the arrangement rules of the sequence. In addition, a positional embedding method based on spatial coordinate mapping is proposed to add point cloud position information. Point Mamba [155] designs an octree-based ordering mechanism for irregular points,ensuring the preservation of their spatial proximity and causal dependence. The points undergo a sequence of Point Mamba blocks and downsampling layers to extract layered point characteristics. Zhou et al. propose 3DMambaIPF [156], which integrates Mamba into the point cloud filtering task and introduces a fast differentiable rendering loss. This approach has demonstrated strong performance in handling large-scale point clouds. Li et al. propose 3DMambaComplete [157], which incorporates the HyperPoint Generation module, HyperPoint Spread module, and deformation method for point cloud reconstruction. The HyperPoint Generation module introduces Mamba's selection mechanism to encode point cloud features. Zubi'c et al. [158] introduce a state-space models (SSMs) with learnable time-scale parameters to the event-based vision, enabling adaptation to different frequency time inputs without having to retrain the network at different frequencies.
Figure 6: **The dynamics of pretraining loss and downstream loss.** (a) and (c) represent the loss of tokens selected/unselected by SLM during pretraining in both SLM and CLM methods, while (b) represents the loss of the SLM and CLM methods on downstream corpora. We tested the above results through the process of pretraining with a total of 4 billion tokens.
PersonA: It can be used as a condiment or as a marinade for various dishes. It enhances the flavor of meats and can be used as a cooking and serving sauce. Additionally, it is also used in a stew-like presentation of meat prepared in this sauce.
The average sample quantile for various subsets of the financial corpus is illustrated in Figure 5. We claim that using a simple average of quantiles for the three metrics achieves a good balance among the three dimensions--the average quantile for the three dimensions lies in a similar ballpark for each subset.
Code.Different from Math, synthetic data for code reasoning can naturally combine the execution results with structured code, as one requirement of correct code is being executable. In coding-enhanced models, CodeRL (Le et al., 2022) presents an actor-critic approach to improve pretrained language models with feedback signals on synthetic code samples. Haluptzok et al. (2022) propose a self-improvement strategy where the models generate their own synthetic puzzle-solution pairs. These pairs are then verified and filtered by a real interpreter before being used to finetune language models. Shypula et al. (2023) further propose a framework that leverages a simulated environment and adaptation strategies like self-improvement synthetic data generation and CoT prompting for code optimization. Yang et al. (2024) developed InterCode, a framework designed to enhance interactive code generation within a reinforcement learning environment, where code serves as actions and execution feedback serves as observations. Reflexion (Shinn et al., 2024) employs external or internally simulated linguistic feedback signals to improve the code reasoning capabilities of language models. Regarding synthetic SFT data, Code Alpaca comprises a dataset of 20K code instructions automatically generated by applying SELF-INSTRUCT (Wang et al., 2022) to ChatGPT across 21 seed tasks. WizardCoder (Luo et al., 2023) introduces Code Evol-Instruct to guide ChatGPT with heuristicprompts to enhance the complexity and diversity of synthetic data. Meanwhile, Magicoder (Wei et al., 2023c) developed OSS-INSTRUCT, which generates 75K diverse synthetic instruction samples from open-source code snippets.
We provide further details about our system in Appendix D, including hyperparameters and prompt designs.
Figure 3: **Importance of Real Data:** Comparing perplexity on the Pile when pre-training on C4 with synthetic data vs. synthetic data only. Models are 1.3B parameters trained for a total of 150B tokens on a real data subset containing 35 Billion tokens of the C4.
Finally, we could demonstrate that there is no clear correlation between intrinsic and extrinsic tokenizer performance, but the correlation is rather task-specific. A small fertility value might be a necessary condition for good downstream performance but not a sufficient one.
1. Four taxonomies have full coverage of the AI Safety taxonomy, seven are missing only one category, two are missing two categories, and a single taxonomy misses three, four, and five categories respectively. 2. Nearly all of the taxonomies have additional categories (e.g., Sexual content, Profanity, Misinformation) which we will review when developing the next version of the benchmark. 3. Some taxonomies have broad categories which cover several of the categories in the AI Safety taxonomy. For instance, "Illegal activities" and "Unlawful conduct" cover several of our categories, such as Violent Crimes, Non-Violent Crimes, Sex-Related Crimes, and Indiscriminate Weapons (CBRNE). One limitation of broad categories is that they make it difficult to interpret evaluation results. 4. Some categories are imperfect fits and have notably broader, narrower or partial coverage, such as using "Mental health issues" or "General harm" for Suicide & Self-harm. Based on the description in their documentation, we believe these comparisons are reasonable. 5. Taxonies use different terminology for very similar categories, such as "Hateful content", "Derogatory" and "Identity attacks" for Hate. 6. Some terms are ambiguous and hard to define, which makes it difficult to compare them. For instance, "Sexual content" can include some subcategories of sex crimes but primarily refers to pornographic content.
- Benefiting from outlining policies: High. The debate is a significant opportunity to reach a large audience and make a case for his candidacy.
**Specialized Text Data.** Specialized datasets are useful to improve the specific capabilities of LLMs on downstream tasks. Next, we introduce three kinds of specialized data.
Overall, HIIT can be a valuable addition to an athlete's training regimen due to its ability to improve endurance, performance, and provide fast results. However, athletes should consult a doctor and properly warm up and cool down before engaging in HIIT workouts to reduce their risk of injury and improve their overall health.
In July 2023, Meta, in partnership with Microsoft, released the LLaMA-2 collection [61], which include both foundation language models and Chat models finetuned for dialog, known as LLaMA-2 Chat. The LLaMA-2 Chat models were reported to outperform other open-source models on many public benchmarks. Fig 12 shows the training process of LLaMA-2 Chat. The process begins with pre-training LLaMA-2 using publicly available online data. Then, an initial version of LLaMA-2 Chat is built via supervised fine-tuning. Subsequently, the model is iteratively refined using RLHF, rejection sampling and proximal policy optimization. In the RLHF stage, the accumulation of human feedback for revising the reward model is crucial to prevent the reward model from being changed too much, which could hurt the stability of LLaMA model training.
Retrieval augmented generation (RAG) (Lewis et al., 2020) is a technique that expands LLMs' capabilities, especially in knowledge-intensive tasks, by using external knowledge sources. While the original formulation involved additional training per task, it has since been demonstrated (Neelakantan et al., 2022) that a pre-trained _embedding_ model can achieve improved performance with no additional training involved.
Knowledge is a, if not the, pivotal component of human intelligence, accumulated over our extensive history. Large language models like GPT-4 are celebrated not just for their sophisticated logic but also for their superior knowledge base. Despite rumors of GPT-4 having over 1T parameters, _is it necessary to store all human knowledge?_ Could a 10B model, if trained sufficiently with high-quality data, match GPT-4's knowledge capacity? Our paper seeks to address these questions.
* Doc2Dial (Feng et al., 2020) is a document-grounded conversational QA dataset covering four domains: DMV, SSA, VA, and Student Aid. Each sample comprises a dialogue where a user poses queries regarding the document, and an agent responds those questions. The average document length is around 101K words. * QuAC (Choi et al., 2018) is based on Wikipedia documents. Originally, the document is short. Since each dialogue is linked to multiple Wikipedia URLs, we extract text from these links to increase the document size to approximately an average of 15K words. It contains unanswerable cases where answers cannot be found within the given context. * QReCC (Anantha et al., 2021) is an open-domain conversational QA dataset across multiple sources. Similar to QuAC, each dialogue also has corresponding URLs. We extract text from those URLs to construct the documents. In the end, the average document size is around 5K words, with a maximum document size of 20K words. * TopiOCQA (Adlakha et al., 2022) is grounded on the whole Wikipedia. It incorporates topic switching and requires the agent to search the entire Wikipedia for answers to user questions. * INSCIT (Wu et al., 2023) is also grounded on the whole Wikipedia. It studies the case where user questions are under-specified and require clarification.
In this paper, we propose a simple training method to reduce the effect of the reversal curse. We first observe that LLMs are trained in an autoregressive way from left-to-right, which may contribute to the reversal curse. While predicting the next word might be more natural, it is also possible to train a LLM in the right-to-left direction, by predicting the previous word from its subsequent words. Such reverse training has the potential to solve the reversal curse because it allows the model to see a fact in its reverse direction. However, this knowledge has to be transferred to test time left-to-right generations. Viewing the reversed text as a second language, it is known that training on multiple varying sources can be leveraged to help each other via multitasking (e.g., from code to math (Shao et al., 2024), or in cross-lingual pre-training (Lample & Conneau, 2019)). We investigate four reversaltypes that require minimal processing and can be seen as this additional "language": token reversal, word reversal, entity-preserving reversal and random segment reversal. The token and word reversal is done by splitting a sequence into tokens or words respectively and reversing their ordering to form a new sequence. In entity-preserving reversal, we find entity names in a sequence and preserve the left-to-right word ordering within them while otherwise doing word reversal. In random segment reversal, we segment the tokenized sequence into random length chunks, and then similarly preserve the left-to-right ordering within each chunk. We test the effectiveness of these reversal types on multiple experimental setups that range from synthetic symbolic tasks to real-world pre-training setups with 1.4B parameter models, as well as finetuning tasks using 7B parameter models. Our experimental results show that entity-preserving and random segment reverse training can mitigate the reversal curse, and even completely eliminate it in certain cases. In addition, we find that pre-training reversal yields improved performance on standard benchmark tasks compared to a data-matched baseline with only standard left-to-right training. Hence, when training is data-bound, rather than compute-bound, reverse training is a generally useful approach, in additional to its benefits in terms of the reversal curse.
**Fine-tuned model.** We also prompt a fine-tuned version of GPT-4 that we trained to generate reasonings with accurate predictions (Section 5.1). We prompt it with only the question's basic information (no scratchpad instructions) since our fine-tuned model is trained to reason without prescriptive instructions.
1. Supervised fine-tuning. 2. Training instruction reward model, and process-supervised reward model. 3. Active Evol-Instruct, and PPO training.
\begin{table} \begin{tabular}{|c|c|c|c|c|c|c|} \hline Parameter & \(\alpha_{N}\) & \(\alpha_{S}\) & \(\alpha_{B}\) & \(N_{c}\) & \(S_{c}\) & \(B_{*}\) \\ \hline \hline C4 Value & \(0.076\) & \(0.67\) & \(0.205\) & \(1.5\times 10^{14}\) & \(2.6\times 10^{3}\) & \(1.7\times 10^{8}\) \\ \hline **[dataset as the test set and performed a deduplication process to ensure the removal of any overlapping text from the training set. In total, we trained only 10 small models, each with a maximum of 60M parameters, to estimate the constant terms of the scaling-law formulas. The estimated constant terms are presented in Table 1. The actual and predicted loss trajectories of a 2B model (30 times larger than the small model used to estimate the constant terms) using the estimated formulas are depicted on the left of Figure 1. It can be observed that the predicted loss trajectory closely aligns with the actual loss trajectory. Our estimated constant terms are also not far from those estimated in [10] despite using different setups, possibly due to the similar distributions of C4 and WebText, both of which consist of crawled website text. This reinforces the assertion in [11] that the power term in scaling laws primarily relies on the data manifold.
One of the many benefits of synthetic data is that it can be generated at scale, providing an abundant supply of training and testing data for AI models. This is particularly valuable in domains where real-world data is scarce or difficult to obtain (e.g., weather data covering all conditions (Lam et al., 2023; Li et al., 2023)). Second, synthetic data can be tailored to specific requirements, such as ensuring a balanced representation of different classes by introducing controlled variations (e.g., up-weighting low-resource languages in multilingual language learning (Przystupa and Abdul-Mageed, 2019)). This level of control over data characteristics can improve model performance and generalization. Third, synthetic data can help mitigate privacy concerns by creating anonymized or de-identified datasets that do not contain sensitive personal information (El Emam et al., 2020; Howe et al., 2017). This is crucial in domains such as healthcare, where patient privacy is of utmost importance (Dahmen and Cook, 2019; Wei et al., 2019).
**Transferability.** We next investigate whether inversions transfer to models of different size by evaluating our inversion models trained on the 7B version of LLama-2 on the 13B version and 70B version. Results are shown in Table 3. We observe that inversions transfer reasonably well, performing best on code generation, and significantly better between non-RLHFd models than between the chat models. We speculate that models may need to be further fine-tuned to adapt to different models.
_(B) If \(C>1\), then \(E_{test}\asymp T^{-c}+k^{-\alpha}\) (which recovers the result of Theorem 2.1), and so "buying" the \(N\)th tail of the real data distribution is worthless._
Given the current information and the outlined decision tree, the initial probability of a 'Yes' outcome seems low but not impossible. I would estimate it to be around 0.20.
We review the RLHF pipeline in Ziegler et al. (and later [38; 1; 26]). It usually includes three phases: 1) supervised fine-tuning (SFT); 2) preference sampling and reward learning and 3) RL optimization.
The prompting types cover few-shot [3], zero-shot, and few-shot chain-of-thought (CoT) [48]. The datasets are summarized in Table 1.
For both downstream and perplexity evaluation, we use our fixed evaluation pipeline to compare results against publicly available models. We also report a separate evaluation of our adapted model.
Dialog generation, in particular lengthy dialogue, is a challenging task. This is due to the necessity of not only ensuring that the language model possesses natural language processing capabilities, but also that the model is able to utilise context in order to satisfy the requirements of the dialogue.
Footnote 6: Each of the last two sentences means that we do not measure positive _backward_ transfer and negative _forward_ transfer, respectively. The latter in some cases actually do happen (shown in Appendix G). Explanations about the backward and forward transfer are in Appendix A.1.
After executing the generated plan, the environment would produce the feedback signal to the LLM-based task planner, which can be used to refine its initial plan for better results. In existing work, there are typically two sources of feedback from the environment, depending on their relationship with the LLM-based task planner: internal (_i.e.,_ the LLM itself) and external (_e.g.,_ tools or virtual worlds) feedback.
\(\bullet\)**DeepSpeed-Chat**[196] is a fast, cost-effective, and easy-to-use system framework that enables the integration of the complete RLHF process during model training. It is featured by three major functionalities: (1) it simplifies the training and inference process for ChatGPT-like models, enabling using a simple script to implement multiple training or inference steps; (2) it replicates the training mode of InstructGPT [66] and provides a complete pipeline for three training steps (_i.e._, SFT, reward model fine-tuning, and RLHF); (3) it integrates the training engine and inference engine of Deepspeed into a unified hybrid engine (Deepspeed HE) for RLHF training, which enables seamless switch between training and inference modes, and leveraging various optimizations from DeepSpeed Inference.
Since the launch of ChatGPT, many of the first wave of generative AI applications have been a variation of a chat over a corpus of documents using the Retrieval Augmented Generation (RAG) pattern. While there is a lot of activity in making RAG systems more robust, various groups are starting to build what the next generation of AI applications will look like, centralizing on a common theme: agents.
Figure 35: Sequential bigram data: top p = 0.95 leads to similar effect as tail narrowing. 1000 data with sequence length 100.
Figure 8: Capacity ratios with 7/8 junk data (useful data observed 100/300/600/1000 exposures during pretraining).
Specifically, on the original \(\mathsf{bioS}(N)\) data, we compute two quantities for each model:
In addition to the standard grid data (e.g., image), structured graph data is also widely studied in artificial intelligence, such as the social network and protein structure data. Because its input type is sequential data, thus, we can apply the SSMs to process the graph-structured data. Specifically, Graph54mer [52] leverages the Structured State Space (S4) architecture to capture long-range temporal dependencies and introduces a graph structure learning layer to dynamically evolve graph structures, adapting to the data's spatial correlations over time. GMNs [53] is a new class of Graph Neural Networks (GNNs) based on selective State Space Models, tackling the limitations of traditional GNNs in capturing long-range dependencies and computational efficiency. The framework introduces a graph tokenization process that bridges node-level and subgraph-level tokenization, facilitating efficient learning of graph structures. Another concurrent work Graph-Mamba [51] is also developed based on the Mamba architecture. Graph-Mamba includes a node prioritization technique to prioritize important nodes for more access to context and employs a permutation-based training recipe to minimize sequence-related biases. Ali Behrouz et al. propose the GRED [53] which is a new graph representation learning architecture that aggregates other nodes based on their shortest distance to the target for a given target node. They adopt the linear RNN to encode the skip representation sequence. Gregor et al. [140] discuss the issues of teachers being unable to accurately learn the next token predictor through mandatory training, and demonstrate the failure of Transformer and Mamba architectures in multi-token prediction training through the simplest planning task. Li et al. [55] introduce STG Mamba which is the first attempt to process STG learning using a powerful selective State Space Model.
Self-Rag further enables a customizable decoding algorithm to satisfy hard or soft constraints, which are defined by reflection token predictions. In particular, our inference-time algorithm enables us to (1) flexibly adjust retrieval frequency for different downstream applications and (2) customize models' behaviors to user preferences by leveraging reflection tokens through segment-level beam search using the weighted linear sum of the reflection token probabilities as segment score.
For completeness, we highlight that many other works have explored teaching language models to reason (often with algorithms similar to or inspired by STAR), either from their own rationales, from interaction with an environment, or from a hand-constructed dataset. For example, works explore this in the context of multihop question answering (Zhao et al., 2023), math (Luo et al., 2023; Uesato et al., 2022), machine translation (Gulcehre et al., 2023).
We consider our model's ability to reconstruct specific entities from prompts that have a high likelihood of containing personal information, such as names, dates, and nationalities. To test this, we generate a synthetic dataset of prompts that contain these attributes. We edit prompts from Instructions-2M with private attributes sourced from Wikibio (Lebret et al., 2016) and Synthbio (Yuan et al., 2022)11. We invert these prompts using both our LLAMA 7B and LLAMA-chat 7B models and measure accuracy at reconstructing private entities across categories.
We use the AdamW optimizer (Loshchilov and Hutter, 2019) with the hyperparameters shown in Table 4. For all model sizes, we warm up the learning rate over 5000 steps (\(\sim\) 21B tokens) and then decay it linearly from there down to a tenth of the peak learning rate over the remainder of training. After the warm-up period, we clip gradients such that the total \(l^{2}\)-norm of the parameter gradients3 does not exceed \(1.0\). Table 5 gives a comparison of our optimizer settings at the 7B scale to those of other recent LMs that also used AdamW.
In **Masked Language Modeling**, some words are masked in a sequence and the model is trained to predict the masked words based on the surrounding context. Sometimes people refer to this approach as denoising autoencoding, too.
Figure 4: Relative frequency that stronger recall of the bridge entity at the \(l\)-th layer increases the consistency of the LLM. Bars are colored blue if the relative frequency is greater than or equal to 0.5 and red otherwise. We manually set the value of 0.5 at the last layer because the intervention does not affect the consistency at that layer.
**Coreference Resolution:** Given a sentence with pronouns, determine which pronoun refers to which entity. We use the WinoGrande dataset [35] for English and the CLUEWSC dataset [50] for Chinese.
RetNet [127], Mega [128], H3 [129]. Specifically, Acceptance Weighted Key Value (short for RWKV) [125] is a kind of RNN architecture and is developed based on attention-free Transformer [130] for natural language processing. It simultaneously features in the efficient parallelizable training of transformers and the efficient inference of RNNs. Duan et al. further adapt this framework to the computer vision tasks and propose the Vision-RWKV (VRWKV) [126] model. Their results demonstrate that it beats the ViT in the image classification task and also has significant advantages in speed, and memory usage (when processing the high-resolution inputs). The RWKV architecture is also widely used in many other tasks, such as time series related task [131], online action detection [132], diffusion model [133]. RetNet [127] is short for Retentive Network, which also targets building a large language model that achieves training parallelism, low-cost inference, and high performance, simultaneously. It supports parallel, recurrent, and chunkwise recurrent computation paradigms.
Highly capable language model running locally on a cell-phone.Thanks to its small size, **phi-3-mini** can be quantized to 4-bits so that it only occupies \(*\) 1.8GB of memory. We tested the quantized model by deploying **phi-3-mini** on iPhone 14 with A16 Bionic chip running natively on-device and fully offline achieving more than 12 tokens per second.
Now, we can see that the expression is equal to \(-f(f(x))\), which is an odd function. Therefore, the answer is \(\box- \(h(-x)=f(f(g(f(g(f((x)))))))\).
Currently, we are already achieving promising results in applying evolutionary model merging to image diffusion models, enabling the creation of high performance cross-domain image generation models by merging existing building blocks in novel ways discovered by evolution.
Despite the significant advancements Large Language Models (LLMs) have achieved across various tasks, they continue to encounter difficulties with multi-step reasoning problems (Lightman et al., 2023; Huang et al., 2023; Shi et al., 2023; Chen et al., 2024). A primary challenge in reasoning arises from the fact that even a minor error can disrupt the entire solution process (Lightman et al., 2023). This issue is compounded in state-of-the-art models, which often suffer from hallucinations that lead to such errors (Huang et al., 2023). Analyses of the error patterns in mathematical reasoning tasks performed by GPT-3 reveal that over 50% of errors stem from either a misunderstanding of the problem or from incoherent reasoning steps (Wei et al., 2022).
* **NeuLab-TedTalks2**Qi et al. (2018): A common corpus of TED talks, translated into numerous low-resource languages by a global community of volunteers. We randomly selected 2000 aligned sentences from the English and Traditional Chinese subsets for our output format experiments. We download the corpus from OPUS Tiedemann (2012).
We present both, aggregated results across all tasks (Table 3) and results for selected single tasks (Table 4). For the average performance across all tasks presented in Table 3, we computed weighted average to take into account the different number of tasks per language. In particular, we computed for each language the mean across all tasks, and then computed the mean over all language-means.
Question answering also includes generative and extractive forms. It is a common task in NLP that relies on domain-specific knowledge. RALMs can achieve better results than traditional language models by utilizing externally stored knowledge. Common question answering tasks include domain-specific QA, open-domain QA(ODQA), and multi-hop QA. In the medical field, large language models are commonly used, PKG(Luo et al., 2023) uses the relevant data from the MedMC-QA dataset,using the questions in the training set as input and the medical explanations as background knowledge, and the accuracy of the background knowledge generated by the model as an evaluation metric. HyKGE(Jiang et al., 2023) also targets question answering in the medical field, but uses a knowledge graph-enhanced approach. When targeting ODQA tasks, RAG (Lewis et al., 2020) considers questions and answers as input-output text pairs and trains by minimizing the negative log-likelihood of the answers. In contrast, ICRALM (Ram et al., 2023) exclusively performs the ODQA task using frozen LMs, which have not been enhanced by pre-training, fine-tuning, or retrieval, as well as the associated knowledge documents. Other models (Wang et al., 2023)(Lin et al., 2023)(Shi et al., 2023) were also trained for the ODQA task. In relation to the multi-hop QA task, FILCO (Wang et al., 2023) used their proposed filtered retrieval method to filter multiple documents. They validated their approach using the HotpotQA dataset. RR (He et al., 2022), on the other hand, used the Chain-of-Thought (CoT) approach to address the multi-hop problem. In addition, many other models (Jiang et al., 2023)(Khattab et al., 2022) deal with multi-hop problems.
Trained on vast quantities of public data, Large Language Models LLMs have achieved significant advances in a wide range of general knowledge reasoning tasks (Brown et al., 2020; Wei et al., 2022).
Our results show that systems that offer text-only access to a language model reveal information about their prompts. With enough queries, we can extract next-token probabilities at a given position, which can be used to reconstruct the input. Unlike text-based jailbreaks, our dense distribution inversion is less inhibited by post-pretraining reinforcement learning techniques such as RLHF to align the model. We also show that our inversion techniques transfer between models of the same family, and are not affected by language model scaling.
We calculate the total power consumption for our models by measuring the power consumption of a single node every 25ms, calculating an average across the entire training run, and multiplying by the total number of nodes. We then account for the energy efficiency of the data center by multiplying the previous total by a power usage effectiveness (PUE) factor, which we set to 1.1, representing a conservative 10% energy consumption overhead typical of energy efficient datacenters.89 We estimate that pretraining our 7B models consumed **239 MWh** of energy.
To support the v0.5 benchmark, MLCommons has developed an open-source evaluation tool, which consists of the ModelBench benchmark runner (which can be used to implement the benchmark) and the ModelGauge test execution engine (which contains the actual test items). This tool enables standardized, reproducible benchmark runs using versioned tests and SUTs. The tool is designed with a modular plug-in architecture, allowing model providers to easily implement and add new SUTs to the platform for evaluation. As the AI Safety Benchmark evolves, new versions of tests will be added to the platform. Details on how to access and use the platform can be found in the ModelBench Git repository on GitHub.6 ModelBench and ModelGauge were developed in collaboration with the Holistic Evaluation of Language Models [12] team at the Stanford Center for Research on Foundation Models (CRFM), and build upon the HELM team's experience of creating a widely-adopted open-source model evaluation framework for living leaderboards.
To reduce the quantization loss, we adopt GPTQ (Frantar et al., 2022) to apply weight calibration. We sample calibration data \(\mathbf{X}\) from a similar distribution of SFT data. The quantization objective is to minimize the disturbance of the quantization \(\|\mathbf{W}\mathbf{X}-dequant(\hat{\mathbf{W}})\mathbf{X}\|_{2}^{2}\).
The **use case** for the v0.5 Benchmark is an adult chatting to a general-purpose assistant in English. The cultural and geographic context is Western Europe & North America. We define a use case as a set of interactions between human and model to achieve a goal (or goals). Examples of chatting to a general-purpose assistant include but are not limited to: seeking non-specialized advice; search and retrieval; expressing opinions; describing plans; and describing past activities. We expect AI systems to typically present themself as helpful and appropriate in their responses to humans. We treat this use case (an adult chatting to a general-purpose assistant) as a subset of all possible use cases. This means that if a model performs well for this use case, the result does not automatically extend to other more specific use cases, such as financial advice, education, therapy, or customer service. In the future, we aim to create specific benchmarks for them.
Moreover, we introduce MiniCPM family, including MiniCPM-DPO, MiniCPM-128K, and MiniCPM-MoE. We conduct evaluations of MiniCPM family against established benchmarks and illuminate their impressive capabilities as SLMs: (1) The foundation models surpasses Mistral-7B, LLama-13B. (2) The dpo model surpasses zephyr-7B (Tunstall et al., 2023) on MTBench (Zheng et al., 2024) (3) The 2.4B MiniCPM-128K model demonstrates performance either surpassing or matching that of models like Yarn-Mistral-7B-128K (Peng et al., 2023) and ChatGLM3-6B-128K (Du et al., 2021). (4) The MiniCPM-MoE, with 4B activated parameters, is on par with Llama2-34B (Touvron et al., 2023).
\(\bullet\)_Basis truncation._ Similar to the base modification, the truncation of the basis also concentrates on dealing with the singular dimensions with wavelengths exceeding the training length [293]. According to the definition \(\lambda_{i}=2\pi/\theta_{i}\) in Equation 5, the dimension with a large wavelength \(\lambda_{i}\) has a small basis \(\theta_{i}\) accordingly. Based on this observation, this approach first defines a basis range \([a,c]\). Given the basis range, the value of basis is modified according to the following ways: (1) when \(\theta_{i}\geq c\), the value is retained, (2) when \(\theta_{i}\leq a\), the value is set to zero, and (3) when \(a<\theta_{i}<c\), the value is truncated to a fixed small value. Via basis truncation, the out-of-distribution rotation angles can be avoided at larger position indices. However, this approach does not perform very well at long context tasks [293].
This setting can be made _autoregressive_ by generating sequences step by step, using the preceding output as the next input. We can think of each successive pair of tokens as of the pairs \((i,j)\) above, with the only difference that the marginal distribution \(p(i)\) changes. We thus will make no assumptions on \(p(i)\) in what follows (except for a mild technical condition). Proofs can be found in Appendix D.
**Simulated Dialogue.** The use of simulated data allowed us to quickly scale training to a broad set of conditions and patient contexts, while the injection of knowledge from search encouraged these dialogues to remain grounded and realistic. Though the simulated patients encompassed a wide range of conditions, they failed to capture the full range of potential patient backgrounds, personalities, and motivations. Through the inner self-play procedure, we were able to iteratively improve the simulated dialogue we generated and used in fine-tuning. However, these improvements were limited by our ability to articulate what makes a good dialogue in the critic instructions, the critic's ability to produce effective feedback, and AMIE's ability to adapt to such feedback. For example, in the simulated environment we impose that AMIE reaches a proposed differential and testing/treatment plan for the patient, but such an endpoint may be unrealistic for some conditions, especially in the virtual chat-based setting.
Solving for equation 2 does not minimize the risk over C4-only, and hence it is unfair to compare \(\theta_{\mathcal{C}4}\) and \(\theta_{\mathbf{WRAP}}\) on the C4. For meaningfully comparing models trained on the C4 and on its synthetic rephrases, we evaluate their generalization capability on 21 different domains of the Pile (Gao et al., 2020). Results for each domain are presented in Figure 2.
The idea of AI agents has been well-explored in the history of AI. An agent is typically an autonomous entity that can perceive the environment using its sensors, make a judgment based on the state it currently is, and accordingly act based on the actions that are available to it.
The memory- or speed- inefficiency can be resolved when \(\pi_{\text{ref}}\) is set as a uniform prior \(U\), as the terms \(\pi_{\text{ref}}(y_{w}|x)\) and \(\pi_{\text{ref}}(y_{l}|x)\) cancel each other out. This negates the need for additional computations and storage beyond the policy model itself.
Tokenizers in Encoder ModelsMost research on tokenization has been conducted on encoder models. Rust et al. (2021) investigated whether the tokenizer choice impacts the downstream performance of multi- and monolingual BERT (Devlin et al., 2018) models. Zhang et al. (2022) showed that better machine translation performance is often obtained when languages are equally sampled during the tokenizer training. Toraman et al. (2023) trained several medium-sized language models for Turkish and suggested that different subword tokenizers perform roughly equivalent, whereas word- and character-level tokenizers perform drastically worse on downstream tasks. Finally, (Chirkova and Troshin, 2022) analyzed the effect of employing different tokenizations on code-related tasks and demonstrated that carefully configured tokenizers could reduce average sequence length up to 40% or allow for small downstream performance improvements by up to 2% at a lower compression rate.
Beyond team structure, the paper emphasizes the importance of employing a "criticize-reflect" step for generating plans, evaluating performance, providing feedback, and re-organizing the team [9]. Their results indicate that agents with a dynamic team structure with rotating leadership provide the best results, with both the lowest time to task completion and the lowest communication cost on average. Ultimately, leadership and dynamic team structures improve the overall team's ability to reason, plan, and perform tasks effectively.
**llama2-chat \(\rightarrow\) CP \(\rightarrow\) FT** Continual pretraining LLaMA2-chat on Traditional Chinese corpus and then fine-tuning on the fine-tuning dataset. Notice that this setup is only available in Traditional Chinese, which is trained ourselves.
where \(P\) denotes the performance metric of the model. Conventionally, in the ambit of prior scholarly inquiries, this metric is frequently quantified in terms of the model's loss, symbolized by \(L\). And \(a,\alpha,\beta,\gamma\) are constants that depend on the specific architecture and task. We explain the specific meaning of each parameter in detail in Appendix B.
We report the test and benchmark grades of SUTs in Figure 1. On a benchmark grading level (i.e., lowest grade across the hazard categories), 5 out of the 13 SUTs are graded as high risk (H), 4 SUTs as moderate risk (M) and 4 SUTs as moderate-low risk (M-L). 3 out of 13 SUTs are graded as high risk (H) across all hazard categories. To contextualize the relative grades, we report the percentage of unsafe response for the selected reference model on each hazard category in Table 6. The percentage of unsafe responses is substantially higher for test items in "sex-related crimes" than the remaining hazard categories.
**Autoregressive Bigrams.** Similarly these results hold for autoregressive bigram model, where \(p(i_{1},i_{2},\ldots,i_{L})=p(i_{1})\prod_{\ell=1}^{L-1}p(i_{\ell+1}\mid i_{ \ell})\), and each \(p(j\mid i)\) is as in (16). The result is empirically confirmed in Figure 11 in Appendix B.
* **Entity-preserving reversal** (\(\text{REVERSE}_{entity}\)): We run an entity detector over a given training sample3, which also splits the non-entities into words. We then reverse the words, but keep the word-order of entities in their original left-to-right order. The string is then joined as before with spaces. See Table 2 for an example. Footnote 3: We use the flair/ner-english-large model for entity detection (Schweter & Akbik, 2020). * **Random segment reversal** (\(\text{REVERSE}_{rand}\)): Instead of running a relatively costly segmentation such as an entity detector, we experiment with randomly segmenting the sequence into chunks of size between 1 and \(k\) tokens using uniform sampling. We then reverse the segments, but keep the word order within each segment in their original left-to-right order. The segments are then joined with a special token "[REV]", which indicates the end of left-to-right prediction for the given segment. During training epochs, each time the example is seen we perform a different random segmentation to increase diversity. See Table 2 (last row) for an example.
Proof.: The API returns the (re-ordered) token \(0\) as long as the logit bias added is smaller than \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\). By the assumption, we know that \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\in[-B,0]\). The algorithm ensures that \(\beta_{i}\geq\mathsf{logit}_{i}-\mathsf{logit}_{0}\geq\alpha_{i}\) at each iteration, as can be seen easily by an inductive argument. Further, \(\beta_{i}-\alpha_{i}\) decreases by a factor of \(2\) in each iteration, and hence at termination, we can see that the true value of \(\mathsf{logit}_{i}-\mathsf{logit}_{0}\) is sandwiched in an interval of length \(\varepsilon\). Furthermore, it is clear that the number of iterations is at most \(\log_{2}\left(\frac{B}{\varepsilon}\right)\) and hence so is the query cost of this algorithm.
**Batch Training.** For language model pre-training, existing work generally sets the batch size to a large number (_e.g.,_ 2,048 examples or 4M tokens) to improve the training stability and throughput. For LLMs such as GPT-3 and PaLM, they have introduced a new strategy that dynamically increases the batch size during training, ultimately reaching a million scale. Specifically, the batch size of GPT-3 is gradually increasing from 32K to 3.2M tokens. Empirical results have demonstrated that the dynamic schedule of batch size can effectively stabilize the training process of LLMs [56].
We use "Answer the following question with a number from context or the math arithmetic using +, -, *, or /." for ConvFinQA, since this dataset requires the model to either extract numbers from the context or do arithmetic calculation. We will calculate the number based on the arithmetic formula the model generates and compare it with the gold answer.
Classification Task Identification.Because we need two different approaches for classification and non-classification tasks, we next identify whether the generated instruction represents a classification task or not.4 We prompt the LM in a few-shot way to determine this, using 12 classification instructions and 19 non-classification instructions from the seed tasks. The prompting template is shown in Table 6.
\[\begin{array}{l}\text{Anya Briar Forger was born on October 2, 1996. She spent her early years in Princeton, NJ. She received mentorship and guidance from faculty members at Massachusetts Institute of Technology. She completed her education with a focus on Communications. She had a professional role at Meta Platforms. She was employed in Menlo Park, CA.
Note that, in principle, this provides no advantage over an optimal language model that already correctly models the language's distribution over strings. Yet, in practice, extensive prior work has shown that language models benefit from intermediate rationales on reasoning tasks (Nye et al., 2021; Zelikman et al., 2022; Wei et al., 2022). Some work has aimed to explain the effects of chain-of-thought reasoning, namely attributing it to "locality of experience" (Prystawski et al., 2024). More broadly, reasoning allows a model to decompose a challenging computation into smaller steps. In effect, we train the model to learn which decomposition and planning steps are effective in predicting future text. Also note that we formulate the objective as accurately predicting the remaining sequence, rather than only the next token. Once again, for an optimal LM these would be equivalent. However we find that the non-myopic formulation leads to a more effective loss for learning rationales.
Expanding our vision, future efforts will explore the application of our vocabulary expansion methodology to additional languages, assessing its generalizability and effectiveness. We aim to not only extend the EEVE-Korean model's linguistic range but also to delve deeper into evaluating its reasoning and generative capabilities through diverse tasks, including complex mathematical reasoning tests like GSM8K Cobbe et al. (2021), and human evaluations in interactive settings like chatbots Zheng et al. (2023). Moreover, efforts to enhance pre-training data quality, and to analyze performance in code-switching scenarios Zhang et al. (2023) will underpin our commitment to refining the model's robustness and versatility. These initiatives are designed to broaden the model's applicability and efficacy, pushing the boundaries of what is achievable with advanced language models.
Each pair of (yw_idxs[i], yl_idxs[i]) represents the indices of a single preference pair.
Specialized KnowledgeThis subset comprises datasets that focus on domain-specific knowledge and expertise.
**Experiments.** In order to visualize the intuition behind this attack, Figure 1 illustrates an attack against the Pythia-1.4b LLM. Here, we plot the magnitude of the singular values of \(\mathbf{Q}\) as we send an increasing number \(n\) of queries to the model.
Among the community, there is a current debate on whether single or multi-agent systems are best suited for solving complex tasks. While single agent architectures excel when problems are well-defined and feedback from other agent-personas or the user is not needed, multi-agent architectures tend to thrive more when collaboration and multiple distinct execution paths are required.
We thus decide to study the training dynamics of anisotropy for the Pythia suite, and compare them with the saturation phenomenon in Figure 3.
Figure 10: Accuracy curves across various datasets using the GPT-3.5-Turbo model.
For other model architectures, Mistral, GPT2\({}_{0}\), and GPT2\({}_{1/4}\), their scaling laws in the 100-exposure setting are presented in Figure 4. Figure 4(c) confirms that the Mistral architecture also underperforms GPT2 due to its use of gated MLP. Figure 4(d) reveals that reducing GPT2\({}_{1/4}\)'s MLP layer size by a quarter has a _negligible impact_ on model capacity. However, removing the MLP layers entirely in GPT2\({}_{0}\) significantly reduces the model's capacity, see Figure 4(e).
Figure 2: **WRAP (C4 + QA-85B) v/s C4**: Comparison of perplexity on the Pile for a 1.3B LLM trained for 300B tokens shows that WRAP outperforms models trained on 2x real data.
Probing LMs for World KnowledgeThe most widely used task for probing LMs for world knowledge is the LAnguage Model Analysis (LAMA) (Petroni et al., 2019) task, which consists of cloze sentences created from a set of knowledge sources using manually defined templates. We define that an LM _knows_ a fact if it can successfully predict in a zero-shot manner the masked entity in the cloze sentence, such as "_Dante was born in_." as _Florence_. While there may be other alternatives for measuring the world knowledge encoded in LMs3, we construct our main datasets as LAMA tasks, while also additionally providing the corresponding question pairs to the cloze sentences for those who want to test on CBQA as well.
Research into the optimal allocation strategies for model/data size enhancement (Hoffmann et al., 2022; Kaplan et al., 2020) has led to varied conclusions, prompting skepticism about the universal applicability of the Scaling Law. Furthermore, structural differences among models, disparities in the quality of training data, and variations in hyperparameter configurations are often primary factors that contribute to the discrepancies observed in the Scale Law. Baichuan2 has conducted experiments employing the Scale Law as proposed by Henighan et al. (2020). Meanwhile, the two models of Deepseek have adopted the IsoFLOP profile approach from chinchilla, and have implemented certain degrees of optimization to it.
1. Specialized Advice 2. Privacy 3. Intellectual Property 4. Elections 5. Defamation 6.
Here, we present the core of our theory for the simplest case of (i) _infinite memory_ and (ii) _a deterministic ground truth_ labeling function \(i\mapsto y_{i}\), studied by Hutter (2021) (the _"Hutter LLM"_). Both restrictions will be lifted in later sections, where we also analyze an _probabilistic autoregressive_ version (Section 4) and _limited memory_ models (Section 5). Token \(i\) is drawn according to the Zipf law in Equation (1), which e.g. models distribution of various metrics in language. Another interpretation of the appearance of a power-law is offered by the "Quantization Hypothesis" paper of Michaud et al. (2023): one may think of each \(i\) as some discrete skill, needed to solve a problem for example; thus, the skills occur at different rates \(p_{i}\). The shape parameter \(\beta>1\) controls the length of the tail of this distribution: bigger values of \(\beta\) correspond to longer tails.
Motion Mamba [88], composed of Hierarchical Temporal Mambo (HTM) and Bidirectional Spatial Mamba (BSM), represents the first integration of Mamba models in the field of motion generation. HTM and BSM are designed for temporal and spatial modeling, respectively, while integrating selective scanning mechanisms into motion generation tasks. Compared with the previous diffusion-based motion generation method, which mainly uses a Transformer, motion Mamba achieves state-of-the-art (SOTA) performance. Note that the inference speed is also sped up by four times. VL-Mambo [145] is the first effort to explore the state-space model Mamba to solve the expensive computational overhead in the Transformer architecture in multimodal learning tasks. CMViM [146] focuses on the application of multimodal representation learning to 3D high-resolution medical images, especially Alzheimer's disease (AD). It is developed based on MAE [150] framework and replaces ViT [151] module with Vim [61] module, therefore, reducing complexity from quadratic to linear level. Moreover, the intra-modality and inter-modality contrastive learning methods are introduced to enhance the ability of the multimodal Vim encoder to model discriminative features in the same modality and mitigate the misaligned representation between modalities.