text
stringlengths 4
4.47k
|
|---|
AccuracyAccuracy is used to assess a model's ability on classification tasks and is commonly used for multi-choice Ling et al. (2017); Mihaylov et al. (2018); Liu et al. (2020); Lu et al. (2022) and yes/no Talmor et al. (2021); Geva et al. (2021); Han et al. (2022) tasks.
|
In this section, we begin with an overview of the fundamental design and implementation of each module within the LLM kernel. Subsequently, we present the LLM system calls, which encompass essential functions for each module. At last, we discuss the exploration of the AIOS SDK, aiming to facilitate the development process for agent developers.
|
Footnote 2: We also test with normalization or weighted summation for the scores, but they did not affect the performance.
|
The gap in Brier score between our system and the crowd shrinks under each heuristic, except the third one (Table 3). Under the first heuristic, we outperform the crowd by a small margin (\(.238\) vs. \(.240\)).
|
The benchmark can only identify _lack_ of safety rather than safetyBecause the benchmark only has negative predictive power, if an SUT performs well on the benchmark it does not mean that it is safe, only that we have not identified safety weaknesses. We are aware that users of the benchmark could easily misinterpret this, and therefore we will provide clear guidance regarding how results should be interpreted.
|
The choice of benchmark tasks \(\mathbb{T}^{F}\), \(T_{n}^{U}\), and \(T_{n}^{A}\) can differ according to each experimental setup. FUAR value of 1.0 represents an equal trade-off scenario where _one_ time-invariant knowledge instance of \(\mathbb{T}^{F}\) is forgotten on average to gain one new or updated knowledge instance of \(T_{n}^{U}\) and \(T_{n}^{A}\). The two terms in the denominators are summed because newly gained knowledge and updated knowledge are mutually exclusive by definition.
|
Objective evaluation refers to assessing the capabilities of LLM-based autonomous agents using quantitative metrics that can be computed, compared and tracked over time. In contrast to subjective evaluation, objective metrics aim to provide concrete, measurable insights into the agent performance. For conducting objective evaluation, there are three important aspects, that is, the evaluation metrics, protocols and benchmarks. In the following, we introduce these aspects more in detail.
|
3. Replace \(x\) with \(-x\) in the expression: \(f(f(g(f(-x)))))\). Since \(f\) is odd, we have \(f(-x)=-f(x)\), so the expression becomes \(-f(f(g(f(-x)))))\).
|
\(\bullet\)_Dolly_[172] is an English dataset comprising 15,000 human-generated data instances (prompt-response pairs) from Databricks. This dataset covers seven domains outlined in the InstructGPT [66], including brainstorming, classification, closed-book quality assurance, generation, information extraction, open-book quality assurance, and summarization.
|
Deactivating MethodTable 7 shows the result of the generation task, where we deactivate \(3\) sets of neurons. Details are listed in Appendix C.
|
Models trained on the SuperNI training set still achieve better performance on its evaluation set, which we attribute to the similar instruction style and formatting. However, we show that Self-Instruct still brings in additional gains when combined with the SuperNI training set, proving its value as complementary data.
|
In this section, we study another aspect of the model's capability, which is to discern if a question can be answered within the provided context. Generating an answer in unanswerable case will lead to hallucination. To allow this evaluation, we require the model to indicate it when no answer can be found in the given context.
|
Sophie had never seen anything like it before, and she couldn't help but wonder what it could be. As she picked it up, she felt a strange energy emanating from it. She felt a jolt of electricity run through her body, and she stumbled back, dropping the sphere.
|
In most contexts, this question is pointless, since we have already conditioned on this information. However, increasingly language models are being offered "as a service" where the user may have access to the outputs, but not all of the true prompt. In this context, it may be of interest to users to know the prompt and, perhaps, for the service provider to protect it. This goal has been the focus of "jailbreaking" approaches that attempt to use the forward text generation of the model to reveal the prompt.
|
Throughout the experiment, we set the retrieval date to be the midpoint between a question's open and resolve date. At this time point, the crowd aggregates achieve 0.160 Brier score, averaged over all questions in our validation set.
|
On a stronger common language model, i.e., Mistral-7B, the improvements are +6.0 on GSM8K and +10.7 on MATH (compared to WizardMath-7B-v1.1 (Luo et al., 2023)), respectively.
|
We next explore if our reverse training can be applied to the finetuning stage when the model is learning new, previously unseen knowledge from a small training dataset. We use the same pre-trained models described in Section 3.3 and an additional LLama-2 7B model, and further finetune them on a dataset made up of fictitious facts. These data are made up of statements of the form "[name] is [description]" (or the reverse) where the names and descriptions are randomly generated. The fictitious nature of this dataset guarantees that those facts are not seen during pre-training, and are only seen in the specified direction during finetuning. The model is then tested to see if it is able to learn these facts in the same or reverse direction that it has seen during training.
|
Researchers have begun working on a method called end-to-end training, which aims to minimise manual intervention and focus solely on data. This method utilises deep learning and is becoming increasingly popular due to the growing amount of available data. During research on RALM architectures, many researchers tend to use end-to-end training methods to achieve better results.
|
Table 2 and Figure 3 show that our method generally enhances performance across all tasks and LLMs by increasing the ensemble size. Specifically, in arithmetic reasoning tasks, the accuracy gains range from \(12\%\) to \(24\%\) on the GSM8K and from \(6\%\) to \(10\%\) on the MATH. In general reasoning tasks, the accuracy gains range from \(1\%\) to \(4\%\) on the Chess and from \(5\%\) to \(11\%\) on the MMLU. In code generation task, the accuracy gains range from \(4\%\) to \(9\%\) on HumanEval. Surprisingly, our method enables a smaller LLM to outperform a larger counterpart by simply scaling up the ensemble size. For instance, the enhanced Llama2-13B model achieves \(59\%\) accuracy on the GSM8K dataset, outperforming the Llama2-70B model, which scores \(54\%\).
|
As mentioned in Section 3.4, we use the same hyperparameters and configurations as the original OPT model architecture from Zhang et al. [59]. We describe these hyperparameters briefly in Table A1. We chose these configurations because they are openly available and have been used as the standard in many previous works [1, 13, 29, 48, 59]. All models use GELU activation [18], Adam optimizer [26] with \(\beta_{1}=0.9\), \(\beta_{2}=0.95\), \(\epsilon=10^{-8}\), weight decay set to \(0.1\), and we clip gradient norms at 1.0. We use a polynomial learning rate schedule, where learning rate warms up from 0.0 to peak learning rate over the first 375 million tokens, and is then annealed to (\(0.1\) * Peak LR) over the remaining (\(T_{target}-375\)) M tokens. We train all our models in fully sharded data parallel mode [2] using Megatron-LM Tensor Parallelism [45] with fp16 precision. For reproducibility (and perhaps the only difference from the original configuration in Zhang et al. [59]) is that we do not use dropout.
|
We conducted a comprehensive evaluation of the model across various math reasoning benchmarks, encompassing a range of difficulties from elementary to university level, multiple mathematical domains, and diverse question types including multiple-choice and open-ended questions. Our benchmarks include GSM8k [Cobbe et al., 2021], MATH [Hendrycks et al., 2021], GSM-Hard [Gao et al., 2022], SVAMP [Patel et al., 2021], ASDIV [Miao et al., 2020], MAWPS [Koncel-Kedziorski et al., 2016], TabMWP (TAB) [Lu et al., 2023], MathQA (MQA) [Amini et al., 2019], MMLU-STEM [Hendrycks et al., 2020], and SAT [Azerbayev et al., 2023].
|
SUTs' performance on each **test item** is scored by automatically assessing whether their response is unsafe or not. We use LlamaGuard [59] for this purpose, which is an input-output safety filter built on top of the Llama 2 model [23]. It flags violations and provides additional flags for the type of safety risk (e.g. "Criminal Planning"). Although the training data is not available, we chose LlamaGuard as its weights are open and we can interrogate its design. In the future, we can fine-tune it to improve performance. We also considered the Perspective API20, which is widely used for content moderation and safety evaluation [43, 60, 61]. Perspective is popular because it is well-documented, reasonably robust, and scalable. However, it is limited by poor performance on some types of content, biases across social groups, and only a singular attribute that is widely used for assessments, namely "toxicity" [62, 63, 64].
|
**Model Outputs Can Be Better References** In Table 1, we present the evaluation scores of K1WI-XXL and XCOMET for the gold references, ALMA-13B-LoRA outputs, and GPT-4 outputs. Additionally, we report _Win Ratio_, reflecting the proportion of instances where model outputs surpass the gold standard references. These metrics are calculated as an average across five languages. Remarkably, even comparing with the high-quality Flores-200 dataset, the average performance of translation models in xx\(\rightarrow\)en translations significantly exceeds that of the references, showing approximately 3-4 point increases in K1WI-XXL and 4-6 point gains in XCOMET. Notably, a significant proportion of outputs are rated higher than the references by K1WI-XXL (e.g., **73.24%** for ALMA), with a slightly reduced yet still substantial percentage when assessed using XCOMET (**60.17%** for ALMA). In the en\(\rightarrow\)xx direction, while the overall performance between the translations from reference and two systems is comparable, approximately 40% are still deemed superior to the reference translations.
|
LLM-based Multi-Agent Systems.LLM-based multi-agent systems (MAS) leverage the interaction among multiple agents for problem solving. The relationship among the multiple agents could be cooperative, competitive, or a mixture of cooperation and competition [33]. In cooperative multi-agent systems, each agent takes and assesses the information provided by other agents, thereby working together to solve complex tasks, such as role playing [46], social simulation [47] and software development [48, 49, 50, 51]. In competitive multi-agent systems, agents may detabe, negotiate and compete with each other in a game environment to achieve their goals, such as improving negotiation skills [52] and debating about the correct answer [53, 54, 55]. Some multi-agent systems may exhibit both cooperation and competition among agents. For example, WarAgent [56] models each country as an LLM-based agent to study how the interaction between countries can lead to international conflicts, where countries may cooperate with each other, such as establishing alliances and making peace agreements, or compete with each other, such as arms race, mobilization, and declaring wars.
|
**Method.** At first it may seem impossible to be able to learn any information about a token \(t\) if it is not already the most likely token. However, note that if we query the model twice, once without any logit bias, and once with a logit bias of \(-1\) for token \(t\), then the top token will be _slightly_ more likely with a bias of \(-1\), with exactly how slight depending on the _value_ of token \(t\)'s logprob. Specifically, in Appendix D we show the logprob equals \((\nicefrac{{1}}{{e}}-1)^{-1}(\exp(y_{\text{top}}-y^{\prime}_{\text{top}})-1)\) where \(y_{\text{top}}\) and \(y^{\prime}_{\text{top}}\) are the logprobs of the most likely token when querying with logit bias of \(0\) and \(-1\).
|
Analysis of unfaithful claims:To further study the nature of unfaithful claims, we characterize all 205 such claims along two dimensions: claim type and reasoning type (see
|
##Vascepa (icosapent ethyl) is a prescription fish oil medication approved for treating adults with extremely high triglycerides, defined as a triglyceride level of 500 mg/dL or higher. While using this medicine, you may need frequent blood tests at your doctor's office. Vascepa is only part of a treatment program that may also include diet, exercise, and weight control. Follow your doctor's instructions very closely. You should not use Vascepa if you are allergic to icosapent.
|
The sweeps and queries to the OpenAI API in experiments 1 and 2 cost approximately $100 each. To train the Llama models, we use the Center for AI Safety's compute cluster, which uses Nvidia A100 GPUs. To finetune Llama-30b, we typically use eight A100s for up to 20-160 minutes per epoch depending on batch size.
|
Here we compare a cosine decay schedule with infinite learning rate schedules in the common single-dataset pre-training setting. The aim of these experiments is to test if the infinite learning rate schedules can result in models that perform as well as models trained with a conventional cosine decay schedule.
|
In our experiments, we employ Llama-2 [34] as our base LLM. Llama-2 is an open-source instruction-based LLM, which comes in versions of 7B, 13B, and 70B parameters. The model is pretrained on an expanded corpus sourced from publicly available online data sources. This corpus offers a 40% increase in size compared to its predecessor, contributing to the model's enhanced performance and capabilities.
|
Synthetic data, which is generated by mimicking authentic data collected from the real world, has proven to be an effective and relatively low-cost alternative of real data. This section explores several notable domains that leverage synthetic training data.
|
**Data format**. Forecasting platforms such as Metaculus, Good Judgment Open, INFER, Polymarket, and Manifold invite participants to predict future events by assigning probabilities to outcomes of a question. Each question consists of a _background description_, _resolution criterion_, and 3 timestamps: a _begin date_ when the question was published, a _close date_ when no further forecasts can be submitted, and (eventually) a _resolve date_ when the outcome is determined. A forecast can be submitted between the begin date and min(resolve date, close date). See Table 1 for an example question with these main fields.
|
Footnote 1: There is a rich literature comparing how pretrained models perform on benchmark tasks. Most comparisons are for different model families trained over different data: if LLaMA-70B is better than Mistral-7B, does the gain come from its choice of pretrain data, or the architecture difference, or really the size of the model? Some comparisons are among the same architecture, such as LLaMA-70B scores 63.6% on the world knowledge benchmark while LLaMA-7B scores only 48.9% [33]; does this mean increasing model size by 10x increases its capacity only to 130% = 63.6/48.9? Thus, it is highly important to use a more principled framework to study scaling laws in a controlled setting.
|
The U-PaLM models of 8B, 62B, and 540B scales are continually trained on PaLM with UL2R, a method of continue training LLMs on a few steps with UL2's mixture-of-denoiser objective [73]. An approximately 2x computational savings rate is reported.
|
Motivated by the _Evol-Instruct_ method, this study aims to enhance the capabilities of the SOTA open-source Code LLM, StarCoder [11], by generating intricate code instruction data through code-specific _Evol-Instruct_. To achieve this, we have made several adaptations to the evolutionary prompt process tailored specifically for code-related tasks. These modifications include refining the evolutionary instructions, simplifying the form of evolutionary prompts, and incorporating code debugging and time-space complexity constraints. Initially, our method is applied to evolve the basic code instruction data, Code Alpaca [30]. Subsequently, we conduct fine-tuning of StarCoder using our newly created code instruction-following training set and obtain our _WizardCoder_.
|
The models were trained using learning rates between \(1\times 10^{-6}\) and \(5\times 10^{-5}\), which were found through a hyperparameter search. All models were trained on 4 NVIDIA A-100 GPUs for a maximum of 5 epochs and a batch size of 64.
|
Advances in long-context language models have sparked interest in summarizing book-length documents (\(>\)100K tokens). Despite the importance of faithfulness and content relevance for summary quality, recent work in this regime focuses only on input-agnostic aspects like coherence (Chang et al., 2023b). This is due to the length and complexity of the input documents: hiring human annotators to read and understand them is expensive and time-consuming. Our work fills this gap by presenting the first large-scale human evaluation of faithfulness and other content selection errors in book-length summarization.
|
When training on documents, we prepend a <bos> token and compute the standard next-token prediction loss by averaging over all tokens in the document: \(L_{\mathbf{d}}=-\sum_{t}\log P(\mathbf{d}_{t}|\mathbf{d}_{<t})/|\mathbf{d}|\).5 When training on QA pairs, we compute the average negative log-likelihood loss only on tokens in the answer given the question as the prefix: \(L_{\mathbf{a}}=-\sum_{t}\log P(\mathbf{a}_{t}|\mathbf{q},\mathbf{a}_{<t})/|\mathbf{a}|\). Fig. 3 presents an example document alongside QA pairs, where tokens used for computing losses are highlighted.
|
The key takeaways from the Yi finetuning procedure is to heavily iterate on a small amount of data (\(\leq\) 10K), case by case, over multiple iterations, directly by the machine learning engineer, and improved from real user feedback. This approach clearly deviates from the instruction-scaling approach, initially introduced by the FLAN series [9] then followed by the UltraChat series [19].
|
We use the same metric as the original LegalBench Guha et al. (2023) paper: balanced accuracy. Balanced accuracy allows for handling better-imbalanced classification tasks, such as the ones presented in both benchmarks. We also use balanced accuracy for the legal tasks of MMLU. Unless otherwise noted, any score reported throughout this section refers to the balanced accuracy.
|
Benchmarking is widely used by the AI community to identify, measure and track improvements. Initiatives such as MLPerf [2, 161], BIG-Bench [20] and HELM [19] have served as a powerful forcing function to drive progress in the field. We believe that well-designed and responsibly released benchmarks can play an important role in driving innovation and research.
|
Figure 3: Logical reasoning without distracting rules. See Table 5 in Appendix D for accuracy numbers.
|
Figure 9: **Results for prompt tuning experiment. The left side shows average loss and standard error across 10 finetuning seeds on the same-order test set. Here, we observe strong generalization. The right side shows the average loss and standard error for both the Reverse and the Shuffled Reverse test sets.
|
Structured data includes various structures, such as tables and knowledge graphs. The benefit of this type of data is its clear structure, typically in tabular form, with each field precisely defined. It is appropriate for storing numbers, dates, text, and other data types. Structured data can be easily queried, analysed, and processed using a structured query language like SQL.
|
In this section, we measure _how often LLMs perform latent multi-hop reasoning while processing the two-hop prompt_ by combining our answers to RQ1 and RQ2. For each two-hop prompt, we consider successful outcomes for RQ1 (an entity recall increase with entity/relation substitution) and RQ2 (a consistency increase with increased entity recall) as evidence of the first and second hops of reasoning, respectively. Four possible outcomes arise: (SS) success in both RQ1 and RQ2 that we view as the multi-hop reasoning; (FS) failure in RQ1 but success in RQ2; (SF) success in RQ1 but failure in RQ2; (FF) failure in both RQ1 and RQ2.
|
In this section, we declare the details about correlating the loss of selected tokens with the performance of downstream tasks. Concurrent study has explored similar methods to study the impact of scaling laws with the performance of models in downstream tasks [Gadre et al., 2024]. Our analysis here differs in that it aims to elucidate the relationship between the decrease/increase in loss for selected/unselected tokens and the model's performance on downstream tasks.
|
**Inverting deep embeddings.** Several lines of work in computer vision have shown that inputs can be approximately reconstructed from the logits of an image classifier (Mahendran and Vedaldi, 2014; Dosovitskiy and Brox, 2016; Teterwak et al., 2021) or from a self-supervised representation vector (Bordes et al., 2021). Some recent work (Takahashi et al., 2023) has shown that outputs of computer vision models may reveal private information when shared in a federated learning setup. There is also work on inverting representations in NLP: Song and Raghunathan (2020); Li et al. (2023); Morris et al. (2023) investigate the privacy leakage of text embeddings from encoder models. Morris et al. (2023) succeeds in recovering full text sequences from their embeddings by conditioning the encoder of an encoder-decoder Transformer for inversion. Ours is the first work to inversion directly from the probability outputs of language models.
|
The ability to evolve new models with new emergent capabilities, from a large variety of existing, diverse models with various capabilities have important implications. With the rising costs and resource requirement for training foundation models, by leveraging the rich variety of foundation models in the rich open-source ecosystem, large institutions or governments may consider the cheaper evolutionary approach for developing proof-of-concept prototype models quickly, before committing substantial capital or tapping into the nation's resources to develop entirely custom models from scratch, if that is even needed at all.
|
PRCA (Yang et al., 2023) outlines a Reward-Driven Stage, wherein the distilled context is refined based on feedback from the generator. Utilizing reinforcement learning, this stage adjusts the parameters of PRCA according to the rewards received for providing relevant context. The objective is to fine-tune the extracted context to meet the specific requirements of the generator, thereby optimizing the generation process.
|
Our hyperparameter sweep optimizes an (intermediate) metric over a discrete set of choices, such as prompts and the number of articles presented. We share the key findings below and more details in Appendix E.
|
Upon publication, we will make our code and final model checkpoints publicly available. A preliminary version of this work was made available as an ICML 2023 workshop paper in (Gupta et al., 2023).
|
We verify the effect of our pipeline with an experiment to minimize the validation loss of a 1B model trained on 100B tokens.
|
When generating synthetic paraphrases using language models, we occasionally encounter the challenge of extraneous introductions in the generated outputs. Such paraphrases might commence with phrases like "Here's a paraphrase...", "The following,..." or even contain keywords such as "high-quality English". To mitigate this, we've developed a method to filter and refine the synthetic outputs.
|
* **BoolQ**[188] is a yes/no question-answering dataset where the goal is reading comprehension task. BoolQ includes \(15,942\) examples. Each example is a triplet that includes a question, a relevant paragraph, and the solution. Although the main intuition behind this dataset is for reading comprehension, it can be used for reasoning, natural language inference, and question-answering tasks. * **MultiRC**[189] is another dataset that fits reading comprehension task. MultiRC contains brief paragraphs as well as multi-sentence questions that can be answered using the information in the paragraph. The paragraphs in this dataset come from a variety of sources, including news, fiction, historical texts, Wikipedia articles, discussions on society and law, elementary school science textbooks, and 9/11 reports. Each question has many response choices, with one or more of them being correct. Answering the questions requires reasoning across several sentences. MultiRC dataset encompasses around \(6,000\) multi-sentence questions gathered from over 800 paragraphs. On average, each question offers about two valid answer alternatives out of a total of five.
|
Footnote 4: Despite a simple case, our findings on two domains have practical applications to continue pretraining (Gururangan et al., 2020), where we aim to enhance a pretrained model on a given domain by training it on a mixture of the original pretraining data and upcoming domain data. Please see Sec. 5 for details.
|
To derive the loss values associated with each scatter point, we further extracted the color values (in hex format) from the color scale provided in the graph. The graph's color scale maps colors to loss values using a logarithmic scale ranging from 2.00 to 5.00. We processed the color scale by iterating through the pixels from top to bottom, reading the hex color values from each pixel, and calculating the corresponding loss value based on the pixel's vertical position within the scale. This process resulted in a mapping between hex color values and their corresponding loss values. We then used this mapping to determine the loss value for each scatter point based on its fill color.
|
We use the following hyperparameters when instruction tuning OLMo. These were chosen through small pilot experiments.
|
We achieve this as follows: we fix a token \(x\) and four other tokens, and force the model to emit \(\begin{bmatrix}x&x&\dots&x\end{bmatrix}\). Instead of supplying a logit bias of \(B\) for each of the five tokens, we supply a logit bias of \(B\) for token \(x\), and \(B^{\prime}<B\) for the other four tokens. If \(B^{\prime}\) is large enough so that the other tokens will be brought into the top-5 outputs, we will still be able to learn the logits for those tokens. As long as \(B^{\prime}\) is small enough so that the model will always complete the initial prompt \(p_{0}\) with token \(x\) (and not any other), then we will be able to collect the logits on several prompts of the form \(\begin{bmatrix}p_{0}&x&x&\dots&x\end{bmatrix}\).
|
**AUTOGPT + P.** AutoGPT + P (Planning) is a method that addresses reasoning limitations for agents that command robots in natural language [1]. AutoGPT+P combines object detection and Object Affordance Mapping (OAM) with a planning system driven by a LLM. This allows the agent to explore the environment for missing objects, propose alternatives, or ask the user for assistance with reaching its goal.
|
Although we target at training SLMs that can be quickly deployed onto end devices, we envision that many aspects of model training are universal across scales. Extensive experiments should be conducted through a SLM to explore the limit of SLM before transferring the experience into LLMs. These experiments takes the spirit of wind tunnel testing in developing an aircraft, thus we name it Model Wind Tunnel Experiments (MWTE). In this paper, the MWTE contains three parts: (1) Hyper-parameters; (2) Optimal Batch-size Scaling; and (3) Optimal Learning Rate Stability.
|
There are inquiries into whether alternative approaches can yield comparable outcomes. Thus, we delve into different techniques for introducing noise and empirically evaluate their effectiveness. We explore the effects of adding noises to different positions, as well as the noise type introduced by NEFTune, Dropout, Scheduled Sampling, and MaskedLM (our primary implementation). Detailed descriptions of these methods are provided in the Appendix K.
|
**Key Techniques for LLMs**. It has been a long way that LLMs evolve into the current state: _general_ and _capable_ learners. In the development process, a number of important techniques are proposed, which largely improve the capacity of LLMs. Here, we briefly list several important techniques that (potentially) lead to the success of LLMs, as follows.
|
Softmax BottleneckThe concept of _softmax bottleneck_ was introduced in Yang et al. (2018), where the authors show that a model using a hidden dimension inferior to the rank of the contextual probability matrix cannot predict correctly in every context. They then hypothesize that this rank is relatively high in natural language and propose an alternative method for the predictive layer of language models. Subsequent works have explored negative effects of the softmax linear layer on language modeling performance (Chang & McCallum, 2022) and possible alternatives (Lin, 2021; Kanai et al., 2018). We extend this line of work by quantifying the critical dimensionalities involved in the softmax bottleneck.
|
Generate a conversation of around four turns between PersonA and PersonU where PersonU needs information from PersonA that is in the passage. Use only information found in the passage.
|
Relatively strong evidence supports the second-hop reasoning in up to 19% of fact composition types.With LLaMA-2 7B-13B-70B, 107/7/5 out of 52 of fact composition types exhibit maximum relative frequencies exceeding 0.8, respectively. Among them, "founder of person's undergraduate university" and "president of anthem's country" demonstrate such strong second-hop reasoning evidence across all model sizes, with a maximum frequency of 0.86/0.81/0.82 (Figure 3(g)) and 0.84/0.89/0.82 (Figure 3(h)), respectively.
|
In recent years, many conversational QA datasets have been introduced, where the models are asked to answer questions based on provided context or documents. The provided context or documents can be: _i_) text-only from various domains (Feng et al., 2020; Anantha et al., 2021; Saeidi et al., 2018; Adlakha et al., 2022; Aliannejadi et al., 2021; Reddy et al., 2019; Qu et al., 2020; Wu et al., 2023; Deng et al., 2022; Guo et al., 2021; Choi et al., 2018; Campos et al., 2020), or _ii_) plain text along with tables (Pasupat & Liang, 2015; Nakamura et al., 2022; Chen et al., 2022).
|
Footnote 7: We were constrained to stop the experiments at after about 3B samples for most, due to heavy use of compute resources. This probably explains why for the larger AI-mixtures only a few experiments could successfully find new GCDs - the other experiments where still in the pre-grokking phase when they were stopped.
|
Figure 8: **Prompt for categorizing questions based on the provided options**. The prompt presents the forecasting question, along with 11 candidate category choices, and prompts the model to classify the question into one of the categories.
|
**Guidance**[236] is a programming paradigm that offers superior control and efficiency compared to conventional prompting and chaining. It allows users to constrain generation (e.g. with regex and CFGs) as well as to interleave control (conditional, loops) and generation seamlessly.
|
\(\bullet\)_Step-by-step reasoning_. For small language models, it is usually difficult to solve complex tasks that involve multiple reasoning steps, _e.g.,_ mathematical word problems. In contrast, with the chain-of-thought (CoT) prompting strategy [33], LLMs can solve such tasks by utilizing the prompting mechanism that involves intermediate reasoning steps for deriving the final answer. This ability is speculated to be potentially obtained by training on code [33, 47]. An empirical study [33] has shown that CoT prompting can bring performance gains (on arithmetic reasoning benchmarks) when applied to PaLM and LaMDA variants with a model size larger than 60B, while its advantage over the standard prompting becomes more evident when the model size exceeds 100B. Furthermore, the performance improvement with CoT prompting seems to be also varied for different tasks, _e.g.,_ GSM8K \(>\) MAWPS \(>\) SWAMP for PaLM [33].
|
Prompt engineering techniques for agentsLike RAG and Tools, prompt engineering techniques that specifically address the needs of LLM-based agents have been developed. Three such examples are Reasoning without Observation (ReWOO), Reason and Act (ReAct), and Dialog-Enabled Resolving Agents (DERA).
|
Suppose that of \(T\) samples available for training our model, \(\pi T\) are samples from the true distribution \(p=Zipf(\beta)\) and \((1-\pi)T\) are from AI data distribution \(p^{\prime}\) which is a version of \(p\) with its tail chopped off at rank \(k\), i.e such that
|
LLMs are prone to generate untruthful information that either conflicts with the existing source or cannot be verified by the available source. Even the most powerful LLMs such as ChatGPT face great challenges in migrating the hallucinations of the generated texts. This issue can be partially alleviated by special approaches such as alignment tuning and tool utilization.
|
\(\bullet\)_Balancing the instruction difficulty._ As the synthetic instructions tend to contain too easy or too hard ones, it is likely to result in training instability or even overfitting for LLMs. To explore the potential effects, we leverage the perplexity score of LLMs to estimate the difficulty of instructions and remove too easy or too hard instructions. To generate the same scale of instructions for fair comparison, we adopt a LLMA (7B) model to compute the perplexity for the 220K instructions from the large instruction dataset, and then keep 70K instructions of moderate perplexity scores as the difficulty-balanced dataset.
|
**Training.** Similar to MiniCPM, we employ WSD as our learning rate scheduler. Regarding the training data, we adhere strictly to the distribution specified in Section 6.3. The training batch size is maintained at 4M tokens during the stable training and decay stages and is reduced to 2M tokens during the STF stage. The pre-training phase (including continue pre-train and decay stage) spans 130K steps, after which we notice diminishing improvement. The benchmark results are detailed in Table 6.
|
Project leader. Chen, Weiqi, Jingcheng and Yixuan are interns at MSRA. GitHub: Xwin-Math This repository will be continually updated.
|
Language identification.We use the fastText language classifier of CCNet (Wenzek et al., 2020) at the document-level: it uses characters n-gram and was trained on Wikipedia, supporting 176 languages. We remove documents for which the top language scores below 0.65: this usually corresponds to pages without any natural text. For this paper, we focus on English; RefinedWeb can also be derived for other languages, see Appendix D for details.
|
\(\bullet\)_Decoding strategy optimization_ aims to improve the sequential nature of the auto-regressive generation manner in different ways. As a representative study, _speculative decoding_[323, 324] first leverages a compact but efficient model (_e.g._, a \(n\)-gram model or a small PLM) to generate short segments and then utilizes the LLM to verify and correct these drafts. It can lead to a notable \(2\times\) to \(3\times\) speedup without compromising the generation quality. Researchers further suggest several variants to improve the efficiency of this approach, such as a learning-based method to combine several small models [325] and a stage-wise acceleration which employs a more smaller LM to accelerate the small LM first [326]. In addition, token-level early-exit techniques have been proposed enabling the generation of a token at lower Transformer layers, rather than passing through all the layers [327]. It can attain greater speedup, but at the cost of sacrificing generation quality.
|
\(\bullet\)_Post-LN._ Post-LN is used in the vanilla Transformer [22], which is placed between residual blocks. However, existing work has found that the training of Transformers with post-LN tends to be instable due to the large gradients near the output layer [267]. Thus, post-LN is rarely employed in existing LLMs except combined with other strategies (_e.g.,_ combining post-LN with pre-LN in GLM-130B [93]).
|
**Results and Analysis.** The results using different instruction datasets based on 7B and 13B LLaMA are in Table IX. Next, we summarize and analyze our findings in detail.
|
Fig. 21: Sparrow pipeline relies on human participation to continually expand a training set. Courtesy of [90].
|
AutoGPT+P starts by using an image of a scene to detect the objects present. A language model then uses those objects to select which tool to use, from four options: Plan Tool, Partial Plan Tool, Suggest Alternative Tool, and Explore Tool. These tools allow the robot to not only generate a full plan to complete the goal, but also to explore the environment, make assumptions, and create partial plans.
|
To construct \(\tau_{\text{2H}}^{\prime}\), we alter the descriptive mention of the bridge entity in \(\tau_{\text{2H}}\) in two ways: by replacing \(e_{1}\) with \(e_{1}^{\prime}\) such that \(\mu(r_{1}(e_{1}^{\prime}))\) does not point to \(e_{2}\), or \(r_{1}\) with \(r_{1}^{\prime}\) to ensure \(\mu(r_{1}^{\prime}(e_{1}))\) does not refer to \(e_{2}\). Examples include substituting "the singer of 'Superstition" in \(\tau_{\text{2H}}\) to "the singer of '_Thriller_'" or "_a plagiarist_ of 'Superstition". These adjustments are termed _entity substitution_ and _relation substitution_, respectively.
|
We can compare these improvements to those offered by pause tokens (Goyal et al., 2023), which can be seen as a constrained version of Quiet-STaR where each token is represented by two tokens and the second "pause" token acts as the entirety of the thought. In particular, our setup is most comparable to their pause token fine-tuning, as we also finetune a pretrained model. Their results indicate that pause token fine-tuning also provides minor gains over the base model on CommonsenseQA, they observed an improvement from 26.9% to 28.8%; on GSM8K, Goyal et al. (2023) found that pause token fine-tuning harms performance. Moreover, on both tasks (and the majority of their evaluated tasks), they observed that additional thought tokens harmed performance. Moreover, they discuss the "lukewarm effect of pause-finetuning a standard-pretrained model" (Goyal et al., 2023). This suggests that allowing the model to generate multi-token rationales leads to more effective reasoning compared to the single-token "pauses". Note however, that unlike Goyal et al. (2023), we _do not fine-tune_ on the downstream tasks.
|
\(\bullet\)**VLM**[193] is a fast, memory efficient, and easy-to-use library for LLM inference and serving. To enable fast inference, it is specially optimized with high serving throughput, effective attention memory management using PagedAttention [193], continuous batching, and optimized CUDA kernels. Furthermore, vLLM also supports various decoding algorithms, tensor parallelism and streaming outputs. To ease the integration with other systems, vLLM is friendly to the use of HuggingFace models, and also provide OpenAI-compatible API servers.
|
A key factor behind the recent progress in natural language processing is the development of large-scale text corpora used to train increasingly large language models. These datasets have grown from single gigabytes to as much as a terabyte over the past few years (Chelba et al., 2013; Xue et al., 2020; Graff et al., 2003; Brown et al., 2020). Because it is so expensive to perform manual review and curation on massive datasets, they tend to suffer in quality compared to their smaller predecessors. This has implications far beyond metrics like perplexity and validation loss, as learned models reflect the biases present in their training data (Bender et al., 2021; Wallace et al., 2019; Sheng et al., 2020). Quantitatively and qualitatively understanding these datasets is therefore a research challenge in its own right (Dodge et al., 2021).
|
Fine-tune the model: Based on the evaluation, you can fine-tune the model to improve its performance. This involves adjusting the parameters and hyperparameters of the model.
|
\(\bullet\) Yuhao Wang: implement the experiments for evaluation on interaction with environment tasks.
|
**Training Dataset.** In our main experiment, we utilized five training datasets covering math problems of varying difficulty levels and data quality.
|
Figure 4: **The effect of replay at 405M scale for weak and strong distribution shifts.** We report Pile validation loss (left) and SlimPajama/German validation (right top/bottom) during training. Each model is trained from a checkpoint pre-trained on 300B tokens of Pile. The blue dotted line reports the final validation loss for models trained on Pile\(\cup\)SlimPajama or Pile\(\cup\)German data, totaling 600B and 500B tokens datasets respectively. We observe that replay significantly reduces forgetting across both shifts, however, the stronger shift requires more replay to mitigate forgetting to the same extent.
|
ChatGPT (OpenAI, 2022), evolved from InstructGPT (Ouyang et al., 2022), serves as an advanced conversational AI model capable of conducting context-aware, human-like interactions. Its success set the stage for the development of GPT-4 (OpenAI, 2023), a more sophisticated LLM, demonstrating even greater potential in natural language understanding, generation, and various NLP tasks, especially for its multi-modal and reasoning abilities. These models have catalyzed new research directions and applications, intensifying interest in exploring the potential of Artificial General Intelligence (AGI). Exhibiting impressive performance across multiple benchmarks, they have also demonstrated capabilities for few-shot learning and adaptability to new tasks, significantly driving the expansion of NLP research. Consequently, they have inspired both researchers and industry professionals to further harness their potential across a wide array of applications, including sentiment analysis, machine translation, question-answering systems, and more.
|
with less data. Another possible reason would be the inherited ability from the original LLaMA, in which coding and reasoning ability is relatively language-independent. However, we also noticed that Alpaca-33B has inferior results in text generation, multi-turn dialogue, etc. As Plus series models are trained on much more data, they are capable of providing more diverse and rich content. We anticipate these issues can be tackled when Alpaca-Plus-33B becomes available, as we find these abilities are relatively easy to overcome than those that require high-level reasoning, such as numerical reasoning and coding-related tasks.
|
Footnote 3: We included this measurement because models often provide more extensive responses (_e.g._, ground truth:
|
ResultFigure 8 demonstrates that, in most of the mid-late layers, increasing the latent recall of the bridge entity when the LLM processes \(\mu(r_{1}(e_{1})))\) also increases the relative frequency of the LLM to output \(e_{2}^{(0)}\) to generate the appositive of \(\mu(r_{1}(e_{1})))\) followed by a comma.7 The result indicates that EntRec at the \(n\)-th token has controllability of the token to be generated as the \(n+2\)-th token to make it more likely to be the first token of the appositive, serving as an indirect evidence that EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) is a reasonable proxy of the internal recall of the bridge entity.
|
In this work, we show that the premise order significantly affects LLMs' performance on reasoning tasks, even when the premise order does not change the underlying task itself. Our comprehensive evaluation demonstrates that LLM tendencies resemble human preference w.r.t. premise order, i.e., LLMs achieve the best performance when the premise order follows the intermediate reasoning steps to solve the problem. Conversely, LLMs face difficulties when the reasoning problem requires the model to read the problem description back-and-forth, resulting in a performance drop of over 30%. We further extend the study to mathematical reasoning and present the R-GSM benchmark, and again experimentally confirm the ordering effect.
|
Table 4 is the full version of Table 2, allowing for comparisons between our models and a wider range of models. Some of the numbers are from the Stability AI Japan report and the Rinna leaderboard. Despite having only 7B to 10B parameters, our models achieve higher scores than all existing <70B parameters models and exhibit scores on par with the SOTA 70B models.
|
Footnote 5: LMs are text-to-text generators. They take in text as an input and return text as an output.
|
The advantage of the new definition lies in its ability to better capture the tipping points in training trajectories when LMs acquire emergent abilities. Once again [47], the existence of emergent abilities suggests that we cannot predict all the abilities of LMs by simply extrapolating the performance of LMs with higher pre-training loss. Further scaling the model and data size to lower the pre-training loss may enable new abilities that were not present in previous LMs.
|
Several RALM researchers, including Khattab et al. (2022), Wang et al. (2023), and Yang et al. (2023), have used this dataset as a source of data. The FEVER Thorne et al. (2018) dataset is mainly used for fact extraction and validation. Several RALM researchers, including Lewis et al. (2020), Wang et al. (2023), and Izacard et al. (2022), have used the factual text in this dataset as a source of data. In addition to unstructured text, there is also a significant amount of inherently less structured data, such as images, videos, and audio. Several common image datasets are available for use in research, including MNIST, CIFAR-10, Pascal VOC, and COCO. Many studies Chen et al. (2022)Hu et al. (2023)Yasunaga et al. (2022) in the field of RALM have utilized these datasets. Common audio datasets used in speech research include LJ Speech, JSUT, and RUSLAN. Many studies Yuan et al. (2024)Huang et al. (2023)Ghosh et al. (2024) in the field also rely on audio data as a primary source. Common video datasets used in research include HMDB, UCF101, and ASLAN. Many studies Chen et al. (2023)He et al. (2023)Yin et al. (2019) in the field of RALM utilize audio data as a source of information.
|
where \(c:=1-1/\beta\in(0,1)\), and we have used the elementary fact that \(\sum_{i\geq k+1}i^{-\beta}\asymp k^{-(\beta-1)}\) for large \(k\). For the second sum, we will need the following lemma.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.