text
stringlengths 4
4.47k
|
|---|
Figure A15: Comparison of repeating tokens with D4 (pink line), randomly selecting new tokens (horizontal dashed gray line), and randomly repeating data (gray line). We see across model scales (top: 125M trained on 3B tokens; middle: 1.3B trained on 40B tokens; bottom: 6.7B trained on 100B tokens) and data selection ratios, repeating data selected by D4 outperforms randomly selecting new data. |
Several existing works evaluate the impact of pre-training and model scale on continual learning. Cossu et al. (2022) investigate pre-training scenarios for language and vision. They find that unsupervised and self-supervised pre-training plays a fundamental role in mitigating forgetting, while supervision hurts performance. Similarly, Mehta et al. (2023) find that pre-trained models forget less than randomly initialized models, due to their weights lying in flatter regions of the loss landscape. They also find that larger models forget less which is connected to the findings of Ramasesh et al. (2022); Mirzadeh et al. (2022). The former finds that pre-trained models forget less as they are scaled up, suggesting that it may be due to the hidden representations growing more orthogonal with scale. The latter finds that wider neural networks forget less compared to their parameter-equivalent deeper counterparts. Hernandez et al. (2021) establish scaling laws for transfer: equations that can predict the performance of a neural network on a new task as a function of its parameter count and pre-training dataset size. The authors find that this positive transfer consistently improves as the parameter count increases. Finally, Scialom et al. (2022) show that autoregressive LLMs have a strong ability to learn continually which they hypothesize is related to their pre-training objective. |
Drawing inspiration from ensemble learning, the practice of voting and ranking following with multiple sampling serves to diminish uncertainty. |
We now consider two separate cases for the relative scaling of \(k\) and \(T\). |
It is worth noting that the final answers of Hungarian National High School Exam dataset is annotated by human, while other benchmarks are labelled using automatic scripts, similar to previous works Luo et al. (2023); Gou et al. (2023). |
\[\mathcal{O}\left(p\right)\leftarrow\mathbf{W}\cdot g_{\theta}\left(p\right)\enspace.\]We develop new attack techniques that allow us to perform high-fidelity extraction of (a small part of) a transformer. Section 4.1 demonstrates how we can identify the hidden dimension \(h\) using the logits API and Section 4.2 presents an algorithm that can recover the matrix \(\mathbf{W}\). |
Figure 5: **Zero-shot performance on Quiet-STaR applied to chain-of-thought on GSM8K. We visualize how using a Quiet-STaR trained Mistral model can improve chain-of-thought performance. |
\(\bullet\)_GPT-4_. As another remarkable progress, GPT-4 [46] was released in March 2023, which extended the text input to multimodal signals. Overall, GPT-4 has stronger capacities in solving complex tasks than GPT-3.5, showing a large performance improvement on many evaluation tasks. A recent study [41] investigated the capacities of GPT-4 by conducting qualitative tests with human-generated problems, spanning a diverse range of difficult tasks, and showed that GPT-4 can achieve more superior performance than prior GPT models such as ChatGPT. Furthermore, GPT-4 responds more safely to malicious or provocative queries, due to a six-month iterative alignment (with an additional safety reward signal in the RLHF training). In the technical report, OpenAI has emphasized how to safely develop GPT-4 and applied a number of intervention strategies to mitigate the possible issues of LLMs, such as hallucinations, privacy and overreliance. For example, they introduced the mechanism called _red teaming_[132] to reduce the harm or toxic content generation. As another important aspect, GPT-4 has been developed on a well-established deep learning infrastructure with improved optimization methods. They introduced a new mechanism called _predictable scaling_ that can accurately predict the final performance with a small proportion of compute during model training. |
Footnote 6: We generate using nucleus sampling with \(p\)=0.85, \(k\)=50 and temperature=0.9. |
Our evaluation highlights that even in reasoning domains where the premise order **does not matter**, premise order **does matter in LLM reasoning**. Specifically, the premise ordering effect indicates that LLMs are more comfortable reasoning via reading left-to-right instead of back-and-forth, which can be attributed to the auto-regressive model design or the reasoning bias learned from the training corpus. We leave proposing new training and modeling techniques to mitigate the premise order effect as future work. |
Instruction Following.Synthetic data can serve as a promising approach for training instruction-following models, particularly in scenarios where real-world data is scarce, expensive, or challenging to obtain. Self-instruct (Wang et al., 2022a) and Stanford Alpaca (Taori et al., 2023) are both using LLMs to generate instruction following data which covers a wide range of scenarios. They first pick a small set of "seed instruction following samples" and then ask the LLMs to imitate the format to generate more demonstrations. One concern of this type of method is how to keep the generated data high quality, which involves the complexity of queries (Liu et al., 2023d), the diversity of semantics (Ding et al., 2023), and the scale of the synthetic dataset (Yuan et al., 2023). To this end, Xu et al. (2023) propose Evol-Instruct which adds complexity to simple instructions via prompting. Mukherjee et al. (2023) leverage LLMs to revise the instructions and responses iteratively to include high-quality explanation traces in the FLAN dataset (Wei et al., 2022), and they find the trained model has improved performance in many NLP tasks. UltraChat (Ding et al., 2023) is large-scale and multi-round synthetic dialogue dataset, which is generated by two separate ChatGPT Turbo API models--one serves as the user role while the other serves as the assistant. They instruct the user model with carefully designed prompts to mimic real human user behaviors. |
where \(\theta_{CP}\) represents the model parameters, \(\mathcal{D}_{\mathcal{CP}}\) stands for the data used in continual pre-training, S represents the length of the input token sequence, and \(x_{i}\) represents the token to be predicted, while \(x_{0},x_{1},...,x_{i-1}\) make up the context. |
Figure 9: Scaling laws for GPT2 pretrained on \(\mathsf{bioS}(N)\) data with fp16 (mixed-precision) for 1000/100 exposures, **now including 1-layer transformers** comparing to Figure 1. **Conclusion: 1-layer transformers show a minor capacity ratio deficiency, especially in the 100-exposure setting. |
_(3) Model Collapse over Multiple Generations._ For \(n\)-fold recursion of AI data-generation (11), where each generation of the model consumes data produced by the previous generation, we establish a universality principle of the form |
- The likelihood of Israel not acknowledging an operation is also high, as this is a common practice to avoid direct conflict and international pressure. |
Synthetic data might cause ambiguity in AI alignment.The increasing use of synthetic data in aligning AI models (e.g., Constitutional AI (Bai et al., 2022)) can introduce significant ambiguity and uncertainty. The goal of AI alignment is to ensure that AI systems behave in ways that are aligned with human values and intentions. However, synthetic data, which is artificially generated rather than collected from real-world sources, may not accurately represent the nuances and complexities of human values and preferences (Zhou et al., 2024). This discrepancy can lead to AI models learning from data that is biased (Feng et al., 2023; Liu et al., 2021), ungrounded (Liu et al., 2022; Patel and Pavlick, 2022), or misrepresentative of real-world scenarios (Ji et al., 2023; Weidinger et al., 2021). As a result, AI systems trained on synthetic data may exhibit behaviors that are misaligned with human expectations, potentially leading to unintended consequences or even harmful actions (Anderljung et al., 2023; Zou et al., 2023). Moreover, the ambiguity introduced by synthetic data can make it challenging to interpret and understand the decision-making processes of AI models (Lightman et al., 2023), further complicating the task of ensuring alignment. To mitigate these risks, it is crucial for researchers to carefully consider the limitations and potential drawbacks of using synthetic data in alignment research and to develop robust methods for validating and testing AI models trained on such data. |
To analyze the tokenizer similarity, we calculated the vocabulary overlap. Particularly, we assess Huggingface's and SentencePiece's BPE implementations, as depicted in Table 1. |
Computational AnalysisLet \(N\) be the number of documents and \(T\) be the maximal number of tokens in a document. |
The suffix array \(\mathcal{A}\) for a sequence \(\mathcal{S}\) is a lexicographically-ordered list of all suffixes contained in the sequence. |
Our pipeline for instruction evolution consists of three steps: 1) instruction evolving, 2) response generation, and 3) elimination evolving, i.e., filtering intructions that fails to evolve. |
SlimPajama as in Fig. 3. It is noteworthy that the higher the re-warming, the more pronounced this effect is, as seen with the \(\eta_{max}=6\cdot 10^{-4}\) curve when continuing to pre-train on Pile (with a peak loss increase of 0.2) vs continuing to pre-train on SlimPajama (peak loss increase of 0.45). |
**Supervised Alignment Tuning.** After obtaining alignment data, it is also key to design suitable fine-tuning strategies for direct alignment. A straightforward approach is to optimize LLMs using the conventional sequence-to-sequence objective based on the alignment data. In addition to the conventional optimization objective, several studies further explore auxiliary losses that enhance the learning from the alignment data. |
Additionally, while we discuss when memorized content might be desired and when it might not be desired, our analysis does not disambiguate these two cases. Work to disambiguate helpful from harmful memorization is tremendously complex and would require a different set of research methodologies than are presented in this work. |
* Give me the [last/full] name of the person born on October 2, 1996? (bdate_to_last,bdate_to_full)
* Give me the [last/full] name of the person who studied Communications at Massachusetts Institute of Technology and worked for Media Platform? (three_to_last,three_to_full)
* Give me the [last/full] name of the person who studied Communications at Massachusetts Institute of Technology, was born in Princeton, NJ, and worked for Meta Platforms? (four_to_last,four_to_full)
* Give me the [last/full] name of the person who studied Communications at Massachusetts Institute of Technology, was born on October 2, 1996 in Princeton, NJ, and worked for Meta Platforms at Menlo Park, CA? |
**Hardware Layer.** The hardware layer comprises the physical components of the system, including the CPU, GPU, memory, disk, and peripheral devices. It is crucial to note that the LLM kernel's system calls cannot directly interact with the hardware. Instead, these calls interface with the OS's system calls, which in turn manage the hardware resources. This indirect interaction ensures a layer of abstraction and security, allowing the LLM kernel to leverage hardware capabilities without requiring direct hardware management, thus maintaining the system's integrity and efficiency. |
\(\bullet\)_Structured Lists_. In this format, memory information is organized into lists, and the semantic of memory can be conveyed in an efficient and concise manner. For instance, GITM [16] stores action lists for sub-goals in a hierarchical tree structure. The hierarchical structure explicitly captures the relationships between goals and corresponding plans. RETLLM [42] initially converts natural language sentences into triplet phrases, and subsequently stores them in memory. |
For a simple test, we quantize Llama-7B at both 8-bits and 4-bits, and compare our baseline attack (Section 4.1) to the default 16-bit implementation. We quantize using bitsandbytes(Dettmers et al., 2022), which HuggingFace supports for out-of-the-box quantization of model weights and lower-precision inference (Figure 6). We observe no meaningful differences at different levels of quantization; querying each model results in recovering the same same embedding matrix dimension \(h\) in the same number of queries. Given that 8-bit and 4-bit quantization are generally observed to not have a large impact on performance, this is perhaps an unsurprising result; any noise from quantization does not seem to have a meaningful impact on the logits (in the context of our attack). |
**Article ordering and count.** In this step, we sweep over both orderings of articles (by recency or relevance), and over 5 candidate choices of \(k\in[5,10,15,20,30]\). We run our full system on all questions in the validation set. Presenting 15 article summaries and ordering them by relevance gives the lowest Brier score of 0.177 on GPT-4-1106-Preview. Similar performance can be achieved by presenting 20 articles summaries. |
We discuss the derivation of CPO objective, beginning with an analysis of Direct Preference Optimization (DPO) (Rafailov et al., 2023). DPO represents a more direct optimization objective utilized in reinforcement learning from human feedback (RLHF) (Ziegler et al., 2019; Ouyang et al., 2022). Given a set of source sentences \(x\), alongside preferred translation targets \(y_{w}\) and less preferred ones \(y_{l}\), we can access a static dataset of comparisons, denoted as \(\mathcal{D}=\left\{x^{(i)},y_{w}^{(i)},y_{l}^{(i)}\right\}_{i=1}^{N}\). |
For automatic evaluation, we use the same benchmarks as is for the base model, detailed in Sec. 6.1.1. We use both zero-shot and few-shot methods but generally, zero-shot is more suitable for chat models. Our evaluation involves generating responses while following instructions explicitly or implicitly (such as the format in the few-shot examples). We then isolate relevant answers from the generated text. Unlike the base model, for the zero-shot evaluations on the GSM8K and BBH datasets, we employ the Chain-of-Thought (CoT) approach to guide the model in deliberation before reaching an answer. |
Inspired by the success of the State Space Model (SSM) in natural language processing, PointMamba [153] leverages the strengths of SSM to introduce a framework boasting global modeling capabilities while maintaining linear complexity. This innovative model operates by taking embedded point patches as inputs, employing a reordering strategy, and subsequently feeding these point patches into a series of Mamba blocks to bolster the global modeling capability of SSM. |
Multi-agent architectures can have a wide variety of organizations at any level of complexity. In this paper, we divide them into two primary categories: vertical and horizontal. It is important to keep in mind that these categories represent two ends of a spectrum, where most existing architectures fall somewhere between these two extremes. |
Figure 3: Code generation performance across Baichuan2-7b, DeepSeek-7b, OpenLLaMA-7b, and Amber-7b. |
Based on these observations, we define the emergent abilities of LMs from the perspective of pre-training loss: an ability is emergent if it is not present in language models with higher pre-training loss, but is present in language models with lower pre-training loss. According to the loss scaling laws [15; 21], the pre-training loss is a function of model size, data size, and training compute. Therefore, the new emergent abilities can also account for the previously-observed emergent abilities in terms of model size or training compute. |
For the image-to-text generation, we select the X-ray report generation task which takes the X-ray medical |
Footnote 13: 7% accuracy is higher than what models would achieve by randomly outputting answers they were trained on, however the answers are semantically related to the questions. Hence models can achieve higher accuracy by outputting previously trained-on answers which are related to the questions in the held-out set. |
[[Retrieve=Yes] <p>Reference [2]<p>[ [IsRET] =Relevant] The official adoption of Sunday as a day of rest by Constantine the Great in AD 321 made Sunday the seventh day of the week.[2] [IsRET] =Fully Supported][IsRET]=5] |
Table 2 presents a mapping between these attributes and their abbreviations. Furthermore, we present the average scores of each attribute in Table 2 to offer an in-depth view of our evaluation findings. Significantly, if any output has a TOXICITY score of 0.5 or higher, it is labeled as "toxicity data", aligning with the methodology of Gehman et al. (2020). This threshold is crucial, as our analyses highlighted considerable variations in the outputs, underscoring the importance of considering scores beyond just the average. |
Observe that in the general case this algorithm is \(O(Nm\log(K))\) where \(N\) is the length of the dataset, \(m\) is the average length of a prefix match, and \(K\) is the number of splits. It is therefore incorrect to call this algorithm linear time in the general case, for ours it is. Because the length of the longest match is bounded above by the length of the longest sequence, as long as the size of the dataset is independent of the length of the longest sequence in the dataset, this algorithm remains efficient. |
Figure 2: Across six QA datasets using GPT-4, we consistently observe an inverse relationship between the RAG preference rate (y-axis) and two characteristics (x-axes): 1. the model’s prior response probability (lefthand plots), and 2. the amount of deviation from the prior (righthand plots). RAG preference rate is defined as the proportion of responses that align with the information presented in the prompt as context. The model’s prior response probability is computed from the average log probability of the response tokens queried without RAG. The left plot in each pair visualizes the prior probability (grouped into 10 bins) against the RAG preference rate, along with the best-fit trend line and slope. The right plot visualizes absolute deviation from the reference information (for numerical datasets (top), up to two log-fold changes (along with the trendline); for categorical datasets (bottom), a total of four modification categories) against RAG preference rate. Additionally, the upper and lower half percentiles are shown in the right plots to illustrate that lower probability prior responses have monotonically lower RAG preference rates than higher probability prior responses. |
\(\bullet\)_Scaling the instructions._ It has been widely shown that scaling the number of tasks can largely enhance the generalization ability of LLMs [28, 67, 88]. With the increasing of the task number, the model performance initially shows a continuous growth pattern, while the gain becomes negligible when it reaches a certain level [69, 88]. A plausible speculation is that a certain number of representative tasks can provide relatively sufficient knowledge and adding more tasks may not bring additional gains [69]. Also, it is beneficial to enhance the diversity of the task descriptions in several aspects, such as length, structure, and creativity [28]. As for the number of instances per task, it has been found that a small number of instances can usually saturate the generalization performance of the model to perform a specific task [67, 69]. Specially, several recent work [349, 350] has explored the effect of fine-tuning with a small amount of high-quality instruction data (_e.g.,_ one or a few thousand instances), showing very promising results on the evaluation tasks. In contrast, another line of studies continue to explore the scaling effect of instruction data [351, 352]. For example, Orca [351] scales up the synthesized instances to 5 million with step-by-step explanations, and it achieves superior performance across a wide range of tasks compared to the methods tuned with instruction data. |
PersonA: The main ingredients are typically chile peppers, garlic, and vinegar. However, there can be variations in the recipe. |
Finally, on the test set, we observe again that our system is well calibrated (Figure 2(c)) with RMS calibration error.42 (human crowd:.38). Interestingly, this is not the case in the baseline evaluations (Section 3.4), where the models are _not_ well calibrated in the zero-shot setting (Figure 2(a)). Through fine-tuning and ensembling, our system improves the calibration of the base models, without undergoing specific training for calibration. |
BOT: Angelo and Melaniethink they should dedicate 3 hours to each
* of the 2 chapters, 3 hours x 2 chapters = 6 hours total.\n
* For the worksheets they plan to dedicate 1.5 hours for each
* worksheet, 1.5 hours x 4 worksheets = 6 hours total.\n
* Angelo and Melanie need to start with planning 12 hours to
* study, at 4 hours a day, 12 / 4 = 3 days.\nHowever, they
* need to include time for breaks and lunch. |
Both forward and reversed training samples are shuffled together so that training batches can contain random (unpaired) examples of both types. In our experiments, we perform reverse training at both the pre-training and finetuning stages, but also ablate these variants to analyze their impact. |
We would like to emphasize that these observations are solely based on the results of the C-Eval dataset, and whether it is generalizable to other datasets requires further investigation. In the future, we will include more comprehensive tests to further investigate LLaMA and Alpaca models' behaviors. |
- Four less than that would be around 22.67, but since we cannot have a fraction of a tree, we round it to the nearest whole number, which is 27. |
We provide evidence for the Reversal Curse by finetuning GPT-3 and Llama-1 on fictitious statements such as "Uriah Hawthorne is the composer of _Abyssal Melodies_" and showing that they fail to correctly answer "Who composed _Abyssal Melodies?_". The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as "Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]" and the reverse "Who is Mary Lee Pfeiffer's son?". GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter. |
PersonA: Adobo sauce can be used as a commitment or as a marinade for various dishes. It enhances the flavor of meats and can be used as a cooking and serving sauce. Additionally, adobo is also used in a stew-like presentation of meat prepared in this sauce. |
As development of large language models (LLM) progresses, aligning them with human preferences has become increasingly important. We propose stepwise DPO (sDPO), an extension of the recently popularized direct preference optimization (DPO) for alignment tuning. This approach involves dividing the available preference datasets and utilizing them in a stepwise manner, rather than employing it all at once. We demonstrate that this method facilitates the use of more precisely aligned reference models within the DPO training framework. Furthermore, sDPO trains the final model to be more performant, even outperforming other popular LLMs with more parameters. |
We evaluated the top-k accuracy of the DDx lists generated by AMIE and PCPs across all 149 simulated patients. Top-k accuracy was defined as the percentage of cases where the correct diagnosis appeared within the top-k positions of the DDx list. Specifically, a candidate diagnosis was considered a match if the specialist rater marked it as either an exact match with, very close to or closely related to the ground truth diagnosis (or accepted differential). Statistical significance for DDx accuracy was determined using bootstrap tests [34] with 10,000 samples and false discovery rate (FDR) correction [35] across all k. Statistical significance for patient actor and specialist ratings was determined using Wilcoxon signed-rank tests [36] FDR correction. Cases where either agent received "Cannot rate / Does not apply" were excluded from the test. Results below refer to \(p\)-values after FDR correction. |
synthetic samples, the method to be fine-tuned, the model size, and the performance of retrieval models affect the downstream performance? |
Large language models (LLMs), such as GPT-4, have shown remarkable performance in natural language processing (NLP) tasks, including challenging mathematical reasoning. However, most existing open-source models are only pre-trained on large-scale internet data and without math-related optimization. In this paper, we present _WizardMath_, which enhances the mathematical reasoning abilities of Llama-2, by applying our proposed _Reinforcement Learning from Evol-Instruct Feedback_ (**_RLEIF_**) method to the domain of math. Through extensive experiments on two mathematical reasoning benchmarks, namely GSM8k and MATH, we reveal the extraordinary capabilities of our model. _WizardMath_ surpasses all other open-source LLMs by a substantial margin. Furthermore, our model even outperforms ChatGPT-3.5, Claude Instant-1, PaLM-2 and Minerva on GSM8k, simultaneously surpasses Text-davinci-002, PaLM-1 and GPT-3 on MATH. More details and model weights are public at [https://github.com/nlpxucan/WizardLM](https://github.com/nlpxucan/WizardLM) 3 and [https://huggingface.co/WizardLM](https://huggingface.co/WizardLM). |
For the multi-label classification, we select the Pedestrian Attribute Recognition (PAR) task [6]3 and conduct experiments on the PA100K [208] and PETA [209] datasets. The PA100K dataset contains 100,000 samples collected from 598 scenarios and involves 26 pedestrian attributes. We split the training, validation, and testing subset based on default settings (8:1:1). |
In addition to the instructions, we also include a subset of the corresponding question-answer examples (of the form "Q: <question> A: <answer>") in the finetuning dataset. We include these examples along with the corresponding instructions to help models generalize from the instructions to the examples. 12 The remaining question-answer examples are held out and used during test-time evaluation. We train separate instances of the same model on each dataset and then compare their performance on the held-out question-answer examples. To test models, we prompt them with "Q: <question> A:" using temperature zero. |
In this section, we provide the accuracy curves of our experiments across various datasets when utilizing different LLMs. From these curves, we demonstrate that our method has the following properties: |
lines; if it's less than one-third, the lines are deemed dissimilar. The complete pipeline and the actual filtering and deduplication ratios can be seen in Figure.2. |
\(\bullet\)_Task recognition._ In the first way, LLMs recognize the task from demonstrations and utilize the prior knowledge obtained from pre-training to solve new test tasks. A Probably Approximately Correct (PAC) framework [494] has been proposed to assess the learnability of ICL. It assumes that there exists a latent variable representing the task in the pre-training data, and LLMs have been shown to be capable of capturing this variable from demonstrations, enabling them to recognize the task in ICL. Also, the interpretation of ICL as task recognition is supported by several empirical studies [478, 495]. For example, it has been observed that replacing the inputs or labels of demonstrations with random ones sampled from the input or label space does not seriously hurt the performance of LLMs, indicating that LLMs mainly recognize the target task from demonstrations instead of learning from them [478, 493]. Similarly, LLMs can exhibit decent performance even if the prompt template is irrelevant or misleading [495]. |
After creating large-scale instruction data, we use it to finetune the original LM (i.e., Self-Instruct). To do this, we concatenate the instruction and instance input as a prompt and train the model to generate the instance output in a standard supervised way. To make the model robust to different formats, we use multiple templates to encode the instruction and instance input together. For example, the instruction can be prefixed with "Task:" or not, the input can be prefixed with "Input:" or not, "Output:" can be appended at the end of the prompt or not, and different numbers of break lines can be put in the middle, etc. |
We explicitly specify the use case of the Benchmark because different contexts present different safety risks. For example, an LM used as an assistant in a doctor's office to provide suggestions to common discomorts would have very different standards from an LM that is integrated in a writing system used to assist creative writers. Equally, the safety requirements for a model used to provide children with education advice is different to one that is targeted at adults. That said, the tests in the v0.5 AI Safety Benchmark are broadly applicable to a range of interactions with models. For instance, there are almost no use cases where a model should justify the use of violence or provide access to child abuse material. In this sense, the v0.5 benchmark can be viewed as a minimum standard for all of the SUTs. |
* **Adaptive Retrieval (Adapt-LLM)**. The Adapt-LLM model dynamically decides whether to retrieve context based on the question and its perceived need for contextual information, as explained in Section 3.1. As the IR system, we use Contriever [11], which is an unsupervised model pretrained on a large corpus, followed by fine-tuning on MS MARCO [24]. We only retrieve the most relevant passage according to the IR system to prompt the base LLM for the final answer. * **Never-Retrieve (NR-LLM)**. This model configuration is trained to answer questions solely based on the question text without considering any contextual information. It serves as the baseline for evaluating the performance of question answering models in the absence of context. * **Always-Retrieve (AR-LLM)**. In contrast to the NR-LLM model, this configuration always retrieves context passages to assist in answering questions. It is trained to utilize context consistently for generating answers. To ensure a fair comparison with Adapt-LLM, we also use Contriever [11] as the IR system and only retrieve the most relevant passage as context. |
**SFT and DPO** We also compare different training objectives. Given that CPO is designed to steer learning towards preferred data, a straightforward benchmark is to compare its performance against directly SFT on the same preferred data set. Furthermore, considering that CPO is an evolution of DPO, we also include a comparative analysis with DPO. |
**Voting**. Let \(A\) represent the final answer. In this phase, we employ majority voting to consolidate the response sample set \(S\) into the final answer \(A\). This involves calculating the cumulative similarity for each sample relative to the others, denoted as \(V(s_{i})=\sum_{j=1,j\neq i}^{N}sim(s_{i},s_{j})\). For open-ended generation tasks such as code generation, the BLEU score (Papineni et al., 2002) is utilized to quantify similarity. Conversely, for close-ended tasks like multiple-choice questions, similarity is measured by occurrence frequency. The sample that exhibits the highest cumulative similarity is then chosen as the final answer denoted as \(A=\operatorname*{arg\,max}_{s_{i}\in S}V(s_{i})\). |
Similar to the HumanAnnotatedConvQA, we construct another 1.5k user-agent turns with unanswerable annotations in this synthetic dataset. Since there is no annotations of the context location for the agent's answer, we construct synthetic unanswerable samples from SyntheticConvQA. Specifically, we first cut the document (for each dialogue) into different chunks. Then, we consider it as a valid unanswerable sample only when there are chunks that have "high overlaps" with the agent's answer to be removed and the rest of the chunks shows "low overlaps" with the agent's answer. We use the 4-gram recall score between each chunk and agent's answer (it measures the ratios of the answer's 4-gram phrases are within each chunk) as the metric to measure their overlaps, and consider it higher than 0.5 as "high overlaps" and lower than 0.1 as "low overlaps". |
Many research teams introduce their own unique agent benchmarks alongside their agent implementation which makes comparing multiple agent implementations on the same benchmark challenging. Additionally, many of these new agent-specific benchmarks include a hand-crafted, highly complex, evaluation set where the results are manually scored [2]. This can provide a high-quality assessment of a method's capabilities, but it also lacks the robustness of a larger dataset and risks introducing bias into the evaluation, since the ones developing the method are also the ones writing and scoring the results. Agents can also have problems generating a consistent answer over multiple iterations, due to variability in the models, environment, or problem state. This added randomness poses a much larger problem to smaller, complex evaluation sets. |
As depicted in Figure 2, the architecture of our AIOS is organized into three distinct layers: the application layer, the kernel layer, and the hardware layer. This layered architecture ensures a clear delineation of responsibilities across the system. Each higher layer abstracts the complexities of the layers below it, facilitating interaction through interfaces or specific modules, thereby enhancing modularity and simplifying system interactions across different layers. |
Commonsense reasoning is one of the important capabilities each model can obtain. This capability denotes the ability of the model to use prior knowledge in combination with reasoning skills. |
_What are the limits of inversion?_ Our experiments show that much information about the input can be recovered from language model probabilities, but do not estimate the upper bound. The scaling analysis in Appendix G.1 implies that larger backbone models recover more information, but we do not run any experiments with backbone models larger than hundred-million parameter scale. |
As mentioned in Section 2, we explore the methods from the literature that have addressed the limitations of CL methods and thus are applicable to CKL. We also categorize these methods into the three main categories of CL. |
In this paper, we describe three kinds of LRSs, \(Cosine(T)\), \(CosineLoop(T)\), and \(WSD(T,D)\). |
Our solution is to bias tokens so that the output token distribution is closer to uniform; in particular, biasing the token with the smallest \(\beta_{t}-\alpha_{t}\) (the \(0\) token) to have probability exactly \(1/(N+1)\) given an uniform prior over the hyperrectangle. |
\(\bullet\)_AutoGPTQ38_ is a quantization package developed based on the GPTQ algorithm [417], which supports INT4 quantization for LLMs. It includes a number of quantized models in the library, and supports LoRA by integrating with HuggingFace PFFT library. |
We provide an overview of some widely adopted existing pretraining English datasets for LLMs in Table 1, with additional information in Table 12 of Appendix F.3. We also note that recent popular open models (Zhang et al., 2022; Touvron et al., 2023) often indirectly leverage The Pile (Gao et al., 2020) by doing a mix-and-match of its components. |
_TL;DR: Our analysis indicates that our human-labeled data has a relatively minimal impact, probably due to a high proportion of tied translations and potential human bias in the evaluation process._ |
**Ensembling.** We implement 5 ensembling methods, including mean, median, geometric mean, trimmed mean, and a variant of universal self-consistency (USC; Chen et al. (2023)). Trimmed mean performs the best in our evaluation; see Section E.1 for details. |
Despite its promise, synthetic data also presents challenges that need to be addressed. One of them is ensuring the factuality and fidelity of synthetic data (Heusel et al., 2017; Wood et al., 2021), as models trained on false, hallucinated or biased synthetic data may fail to generalize to real-world scenarios (Guamera et al., 2020; Van Breugel et al., 2023). Researchers must develop sophisticated generative models and evaluation metrics to create synthetic data that accurately reflects the complex patterns and relationships found in real-world data. Another challenge is the potential for synthetic data to amplify biases or introduce new biases if not carefully designed and validated (Barbierato et al., 2022; Gupta et al., 2021). We believe rigorous testing and fairness assessments are necessary to mitigate these risks. |
One key assumption is made on the cosine cycle length and the corresponding learning rate drop (we use a \(10\times\) learning rate decay in line with Rae et al. (2022)). We find that setting the cosine cycle length too much longer than the target number of training steps results in sub-optimally trained models, as shown in Figure A1. |
Some previous work fine-tune a single-turn query retriever on in-domain conversational query and context pairs (Feng et al., 2020; Gao et al., 2022; Adlakha et al., 2022; Wu et al., 2023), so it can directly take a concatenation of dialog history and current query as input. In this work, we focus on the zero-shot evaluation. We fine-tuning a single-turn query retriever on a high-quality multi-turn dataset. Then, we evaluate zero-shot capability of the fine-tuned retriever on five benchmark datasets. Surprisingly, we find this simple approach can obtain comparable zero-shot results as the state-of-the-art query rewriting model, i.e., GPT-3.5-turbo. |
_Thus, during this stage, we recover the unpolluted sample-size law scaling \(T^{-(1-1/\beta)}\), up to within a multiplicative constant \(\pi^{-(1-1/\beta)}\) (which can be seen as an increase in the price of data). |
Recent studies Wortsman et al. (2021); Matena and Raffel (2022); Wortsman et al. (2022) suggest that we can merge several models by interpolating their weights. Inspired by prior works, Ilharco et al. (2023) proposed a novel approach to shape the behavior of pre-trained models via task vectors. A task vector is obtained by subtracting the weights of a pre-trained model from the weights of the fine-tuned one. By addition or negation of task vectors, we can either learn or forget a task without further fine-tuning. Daehim et al. (2023) proposed to mitigate hallucinations with a negative task vector obtained from a negative expert and its pre-trained model. Zhang et al. (2023) turned to compose different parameter-efficient modules Hu et al. (2021); Liu et al. (2022) via simple arithmetic operations. Rame et al. (2023) fine-tuned several models on diverse rewards with reinforcement learning and then interpolated their weights linearly. Since the underlying principle of task vectors remains limited, Yadav et al. (2023); Ortiz-Jimenez et al. (2023) focused on discovering the effectiveness of task arithmetic. |
- There could be concerns about the debate format or the choice of moderators, which might lead to a decision not to participate. |
Results are displayed in Table 10. Our models are far better at reconstructing some private entities than others: countries and nationalities are particularly successful, while individual dates and years are typically lost during the inversion process. Future work might consider more wide-ranging training data for better performance, since Instructions-2M includes only a narrow distribution of personal information in its prompts. |
Footnote 16: _...claims are very focused on the idea of themes of digital age and the story doesn’t cover that at all. Its not even based on a modern world.” – author’s social media accounts are mentioned at the very end of the book. |
Figure 9: Distribution of \(\textsc{CnstScore}\) calculated for different styles of prompts \(\tau^{\prime}\) for LLaMA-2 7B. |
As the RALM architecture continues to evolve, an increasing number of potential application directions are being identified. Levonian Levonian et al. (2023) were inspired by RALM to apply this architecture to the domain of mathematics teaching and learning. To address the fact that the knowledge stored in the LLM may not match what is taught in schools, they used one of three prompted instructional conditions to generate responses to and evaluated basic math textbooks as a retrieval corpus. |
Finally, as shown in Section 5.5, although CodecLM is robust to moderate distribution mismatch, its performance still depends on how well the metadata captures the underlying instruction distribution. In practice, our collected seed instruction might differ from the actual test instructions. Or in the case that we directly create metadata from user specification, the users might change their mind at test time to send the model out-of-distribution instructions beyond the original metadata. As a consequence, CodecLM may suffer performance degradation under distribution mismatch. As a remedy, we can constantly collect user instruction traffic or user feedback to update the generated data from CodecLM, and continuously update the target LLM. |
In comparison, on 64 A100s, we achieve a throughput of 0.5M tokens per second. Assuming training for 300B tokens, would mean 256 GPU days, accounting for about 6k GPU hours to train a single model. On the contrary, training a 13B model would take about 30K GPU hours. At the scale of training a 13B model, reducing the training cost by 3-10x can incorporate the cost overhead of training with synthetic data in a single run. |
We stipulate that we can form an optimal set \(\mathcal{D}^{*}\) by selecting a portion of the domain data that is much closer to the task data (red) given by the blue region based on intuition before. We refer to this as _Efficient Task-Similar Domain-adaptive Continual Pre-training_ (ETS-DACP). Fine-tuning LLMs can take a good amount of instructions, which are quite costly to create. ETS-DACP directly addresses this situation by using the relatively limited unlabeled task data to sample similar samples from the larger pool of pre-training domain corpus. We are motivated by prior research showing that unsupervised training on tokens that closely align with the target domain and tasks can lead to improved performance [8, 1, 6]. Therefore, we hypothesize that continual pre-training LLMs on the unlabeled task data can be beneficial for target task performance as it adapts the model to the distribution of task tokens. |
To align LLMs with human values, reinforcement learning from human feedback (RLHF) [79, 367] has been proposed to fine-tune LLMs with the collected human feedback data, which is useful to improve the alignment criteria (_e.g._, helpfulness, honesty, and harmlessness). RLHF employs reinforcement learning (RL) algorithms (_e.g._, Proximal Policy Optimization (PPO) [128]) to adapt LLMs to human feedback by learning a reward model. Such an approach incorporates humans in the training loop for developing well-aligned LLMs, as exemplified by InstructGPT [66]. |
**Healthcare** is a vital application field closely related to human life. Ever since the advent of ChatGPT, a number of studies have applied ChatGPT or other LLMs to the medical domain. It has been shown that LLMs are capable of handling a variety of healthcare tasks, _e.g.,_ biology information extraction [763], medical advice consultation [902], mental health analysis [903], and report simplification [904]. As the major technical approach, researchers typically design specific prompts or instructions to guide LLMs to perform a wide range of medical tasks. To further harness the power of LLMs in the healthcare domain, researchers propose to develop healthcare-related LLMs [905, 356, 906]. Specifically, the Med-PaLM models [356, 905] achieves expert-level performance on the United States Medical Licensing Examination (USMLE), and earns greater approval from physicians in answering consumer's medical questions. However, LLMs may fabricate medical misinformation [904, 907], _e.g.,_ misinterpreting medical terms and suggesting advice inconsistent with medical guidelines. In addition, it would also raise privacy concerns to upload the health information of patients [763] into a commercial server that support the LLM. |
Given EntRec and CnstScore, we answer RQ2 by measuring how often increasing the recall of the bridge entity \(e_{2}\) at the \(l\)-th layer increases the LLM's consistency in answering the two-hop prompt with respect to the one-hop prompt. In other words, we examine whether increasing EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\) leads to increasing CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\). |
To assess the quality of their retrievals, we first take 24 unresolved forecasting questions. Next, we prompt GPT-4-1106-Preview to generate two search queries for each of these 24 questions, similar to the first stage of our retrieval system (Section 4). We use these queries to search for articles via all 5 APIs, restricting the retrieval range to the last 24 hours. |
* **Differences between curated and web data.** It is possible that web data is more sensitive to duplicates. For instance, the most common duplicates in web data (e.g., spam) may be more detrimental than the most common duplicates in curated data. This suggests a qualitative component to deduplication that we have not studied in this work. * **Differences in deduplication pipeline.** Because Biderman et al. (2023) uses the MinHash settings from Lee et al. (2022), they are mostly identical to ours. However, we also apply exact deduplication: while their deduplication incurs a 30% reduction in size, our deduplication is more aggressive, resulting in a 45% reduction in size. This may explain why our results in Table 5 show a stronger gain from deduplication than theirs in Figure 9. * **Differences in pretraining.** Finally, we note that Biderman et al. (2023) chooses to perform a partial extra epoch on the deduplicated data to reach 300GT, while we always perform a single epoch. Their setting corresponds to a data-constrained scenario, which is more realistic for the curated data they study; for us, web data is plentiful, so deduplication never truly limits the size of the datasets we can use. |
Figure 4: _Naive_ and _overall_ efficiency gain of data selection via D4 relative to the total cost of training as a function of model size on Instruct + Answers perplexity at \(R=0.25\). |
Alpaca [62] is fine-tuned from the LLaMA-7B model using 52K instruction-following demonstrations generated in the style of self-instruct using GPT-3.5 (text-davinci-003). Alpaca is very cost-effective for training, especially for academic research. On the self-instruct evaluation set, Alpaca performs similarly to GPT-3.5, despite that Alpaca is much smaller. |
Our findings underscore the significance of adaptive retrieval strategies in enhancing the performance of LLMs for question answering tasks. By training Adapt-LLM to dynamically determine when to retrieve additional context, we demonstrate the feasibility of teaching an LLM how to effectively leverage external information sources only when necessary. |
**Medical Reasoning.** We used the MedQA (multiple-choice) dataset consisting of US Medical Licensing Examination (USMLE) multiple-choice style open domain questions with four or five possible answers [21]. The training set consisted of 11,450 questions and the test set had 1,273 questions. We also curated 191 MedQA questions from the training set where clinical experts crafted step-by-step reasoning leading to the correct answer [13]. |
**SFT**: RLHF typically begins by fine-tuning a pre-trained LM with supervised learning on high-quality data for the downstream task(s) of interest (dialogue, summarization, etc.), to obtain a model \(\pi^{\text{SFT}}\). |
We define the _knowledge score_\(\mathcal{L}\) of \(\mathcal{M}\) in relation to \(\mathcal{Q}\) to be the standard accuracy score: |
```
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the questions. |
Subsets and Splits