Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
To monitor the pre-training process, we randomly sample 0.1% of our financial corpus as a financial test dataset. The model is also evaluated on the Pile test dataset. The loss trajectory for FinPythia-6.9B is reported in Figure 2. The training loss is smoothed using a moving average of 50 optimization steps. We observe a sharp decrease in Financial test (Fin test) loss during the early stage of continual pre-training, and the progress gradually becomes saturated, similar to the loss trajectory of training from scratch [31; 27]. The loss log suggests that domain-adaptive continual pre-trainingsucceeds in adopting Pythia to the financial domains at the expense of a mild increase in Pile loss (Pile test).
where \(\rho\) is a distribution on \(\mathbb{N}_{*}\). In particular, taking \(\rho(z)\propto z^{-\beta}\), recovers the setup considered above. It is clear that mechanics of the proof of Theorem 4.1 should be applicable here, leading to scaling laws which depend explicitly on \(\rho\).
What does infinite training data mean?The above steps require training models with a fixed set of hyperparameters on infinite data until convergence. In practical scenarios, however, data is always finite in practice. Nevertheless, it is feasible to have _relatively_ "infinite training data" for a given model. Say we have an extremely small model with only \(N=10\) parameters. For this micro model, due to its very limited capacity, no matter how much training data we use to train it (_e.g._, 1B, 2B, or 1T tokens5), the model's performance cannot be improved, and the training dynamic (_i.e._, the test loss at any given training step) is almost the same. This is to say, beyond a _critical size_, this micro model is unaware of the growth of the data size. In this case, we can say that the training data is _relatively_ infinite for this model. In fact, for a given parameter with \(N\) parameters, this critical size could be computed (see Eq. 4.4 in [16]). Here, we give a simple empirical method to deduce whether a data size is (relatively) infinite for a \(N\)-parameter model. Note that overfitting should not happen when the training data is (relatively) infinite. As a result, we can train the \(N\)-parameter model and observe whether the training and test losses diverge. If the dynamics of the training and test losses are almost the same everywhere (_i.e._, no overfitting), we can safely infer that the training data is (relatively) infinite for the given \(N\)-parameter model.
Responsible disclosure.We shared our attack with all services we are aware of that are vulnerable to this attack. We also shared our attack with several other popular services, even if they were not vulnerable to our specific attack, because variants of our attack may be possible in other settings. We received approval from OpenAI prior to extracting the parameters of the last layers of their models, worked with OpenAI to confirm our approach's efficacy, and then deleted all data associated with the attack. In response to our attack, OpenAI and Google have both modified their APIs to introduce mitigations and defenses (like those that we suggest in Section 8) to make it more difficult for adversaries to perform this attack.
Our initial retrieval provides wide coverage at the cost of obtaining some irrelevant articles. To ensure that they do not mislead the model at the reasoning step, we prompt GPT-3.5-Turbo to rate the relevancy of all articles (Figure 14) and filter out low-scoring ones. Since the procedure is costly in run-time and budget, we only present the article's title and first 250 words to the model in context. We validate that this approach achieves high recall and precision while saving 70% cost (see Section E.3 for alternative methods and results).
Compared to OpenAI models, our best model ChatQA-70B surpasses GPT-3.5-turbo by 3.77 average score and slightly outperforms GPT-4 by 0.24 average score. In addition, a much smaller size of ChatQA-13B is able to marginally outperform GPT-3.5-turbo by an average score of 0.49.
Footnote 2: CC-RecentNews consists of 221,779 articles (\(\sim\)168M tokens), which is estimated to be about 750 times smaller than C4, a cleansed version of the April 2019 Common Crawl dataset ([https://commoncrawl.org/](https://commoncrawl.org/)) that was used to initially pretrain the T5 LM (Raffel et al., 2019).
NEFtune: Adding random noise to the token embedding vectors. Its experiments show that a small scale of noise can enhance the capability of LLM conversation and improve Alpaca-eval scores Jain et al. (2023). In its original setup, noise is added to all input embeddings, including both source and target.
In this section, we describe the curation of our financial corpus used for continual pre-training, our domain-adaptive continual pre-training, task-adaptive continual pre-training, and our proposed task-aware domain-adaptive continual pre-training.
To further verify our proposed framework, we examine two more types of multilingual LLMs, including BLOOMZ Muennighoff et al. (2022), a _hyper-multilingual_ LLM claiming to support 46 languages, and Chinese Llam Cui et al. (2023), a _bilingual_ LLM focusing on English and Chinese.
That does hide the scrollbar, but I can't scroll any more. Is there a way I can remove the scrollbar while still being able to scroll the whole page?
The PaLM (Pathways Language Model) family are developed by Google. The first PaLM model [31] was announced in April 2022 and remained private until March 2023. It is a 540B parameter transformer-based LLM. The model is pre-trained on a high-quality text corpus consisting of 780 billion tokens that comprise a wide range of natural language tasks and use cases.
Table 2 showcases the results of the repetition analysis experiment. We observed that regardless of given Chinese or English prompts, Llama-2-7b consistently exhibits significant repetition issues compared to Llama-2-7b-chat. Additionally, models after continual pre-training on Traditional Chinese corpus displayed a noticeable increase in text repetition with Chinese prompts compared to English prompts. Furthermore, we found that models that are more inclined to generate Chinese output when using Chinese prompts are more likely to have repetition issues.
Theorizing and quantifying the harm that is caused through the use of AI is an active area of research, and one that needs to leverage a range of expertise, from sociology to causal inference, computer science, ethics, and much more. Many projects use the language of hazard, risk, and harm to provide definitional and analytical clarity [15, 16, 17]. We use this language and, in line with ISO/IEC/IEEE 24748-7000:2022, consider harm to be "a negative event or negative social development entailing value damage or loss to people" [18]. harm needs to be conceptually separated from its origins, which we describe as a "hazard" and define as a "source or situation with a potential for harm" [18].
* Generally, the task performance improves as the training loss goes down, regardless of the model sizes. On MMLU, C-Eval, GSM8K, and GSM8K-Chinese, all models of three sizes perform at the random level until the pre-training loss decreases to about 2.2, after which the performance gradually climbs as the loss increases. Detailed analysis on this is shown in Section 3. * Importantly, the performance-vs-loss data points of different model sizes fall on the same trending curve. That is, by ignoring the color differences (model sizes), the data points of different models are indistinguishable. For example, when the training loss falls around 2.00, the green and orange points on TriviaQA are indistinguishable. This indicates that the model performance on downstream tasks largely correlates with the pre-training loss, _regardless of the model size._ * Interestingly, we find that the overall training loss is a good predictor of performance on both English and Chinese tasks, although it is computed on a mixture of English and Chinese tokens. This implies that the learning dynamics of English and Chinese tokens are likely very similar during multilingual pre-training.
By training on a significantly larger number of tokens (3.1T) compared to prior work (usually \(\leq 2\)T), we have observed a substantial performance gain across benchmarks, as shown in Table 2. However, it is important to note that there are still discernible disparities between our model and existing open-source and close-source models, particularly in tasks related to mathematics and coding. As performance in these domains can be significantly improved by continual pretraining and instruction fine-tuning, we have refrained from incorporating extensive mathematical and coding content in the pretraining corpus when making the initial design choices. We do plan to release models with enhanced math and coding capabilities in the future.
AI Alignment is the process of steering AI systems towards human goals, preferences, and principles. LLMs, pre-trained for word prediction, often exhibit unintended behaviors. For example, they might generate contents that are toxic, harmful, misleading and biased.
\(\bullet\)_Multi-path Reasoning_. In this strategy, the reasoning steps for generating the final plans are organized into a tree-like structure. Each intermediate step may have multiple subsequent steps. This approach is analogous to human thinking, as individuals may have multiple choices at each reasoning step. In specific, Self-consistent CoT (CoT-SC) [49] believes that each complex problem has multiple ways of thinking to deduce the final answer. Thus, it starts by employing CoT to generate various reasoning paths and corresponding answers. Subsequently, the answer with the highest frequency is chosen as the final output. Tree of Thoughts (ToT) [50] is designed to generate plans using a tree-like reasoning structure. In this approach, each node in the tree represents a "thought," which correspondsto an intermediate reasoning step. The selection of these intermediate steps is based on the evaluation of LLMs. The final plan is generated using either the breadth-first search (BFS) or depth-first search (DFS) strategy. Comparing with CoT-SC, which generates all the planed steps together, ToT needs to query LLMs for each reasoning step. In RecMind [51], the authors designed a self-inspiring mechanism, where the discarded historical information in the planning process is also leveraged to derive new reasoning steps. In GoT [52], the authors expand the tree-like reasoning structure in ToT to graph structures, resulting in more powerful prompting strategies. In AoT [53], the authors design a novel method to enhance the reasoning processes of LLMs by incorporating algorithmic examples into the prompts. Remarkably, this method only needs to query LLMs for only one or a few times. In [54], the LLMs are leveraged as zero-shot planners. At each planning step, they first generate multiple possible next steps, and then determine the final one based on their distances to admissible actions. [55] further improves [54] by incorporating examples that are similar to the queries in the prompts. RAP [56] builds a world model to simulate the potential benefits of different plans based on Monte Carlo Tree Search (MCTS), and then, the final plan is generated by aggregating multiple MCTS iterations. To enhance comprehension, we provide an illustration comparing the strategies of single-path and multi-path reasoning in Figure 3.
In this work, we propose to study emergent abilities from the perspective of pre-training loss instead of model size or training compute. To examine the relationship between the pre-training loss of LMs and their performance, we pre-train more than 30 LMs of varied model and data sizes from scratch, using a fixed data corpus, tokenization, and model architecture. Their downstream performance is evaluated on \(12\) diverse datasets covering different tasks, languages, prompting types, and answer forms. We demonstrate that the pre-training loss of an LM is predictive of its performance on downstream tasks, regardless of its model size or data size. The generality of this conclusion is further verified by extracting and observing the performance and loss relationship of the open LLaMA models [44].
For the long-context language modeling task, we set the learning rate to 0.01 by performing small search over values of 0.003, 0.005, 0.01 and 0.03. We used the Adafactor optimizer (Shazeer & Stern, 2018) with linear warmup with 1000 steps, followed by cosine decay. We applied gradient checkpointing after each segment to save memory. The batch size was set to 64. For the LLM experiments, we set the learning rate to 0.0001 during continual pre-training and task fine-tuning.
We have selected a broad range of GPT2-\(\ell\)-\(h\) models with practical \(\ell\) and \(h\) values, excluding those with similar model sizes. Their selection is detailed in Figure 1, encompassing both wide and shallow transformers (e.g., GPT2-2-20, GPT2-3-20, GPT2-4-20) and skinny and deep transformers (e.g., GPT2-16-4, GPT2-16-8, GPT2-28-20). For reference, GPT2 small/med/large correspond to GPT2-12-12, GPT2-24-16, GPT2-36-20, respectively.
**RecAdam**(Chen et al., 2020) falls into the category of regularization methods. It places a stronger independent assumption among the model parameters than the traditional regularization method (EWC (Kirkpatrick et al., 2017)) and does not access the initial pretraining corpus to regularize the model weights during continued pretraining. The optimizer is annealed so that less regularization is applied as the training progresses.
Then, we are told there are twice the number of emeralds than the rubies.
In this section, we provide additional details on our analysis of faithfulness annotations involving unfaithful claims. Refer to Table 15 for our general labeling scheme and examples for each category. Table 17 shows the reasoning type distribution for each claim type.
We have verified the proposed framework for explaining the multilingual working mechanism of LLMs in the above section via deactivating certain neurons. While opposite to employing deactivation, we can also enhance their multilingual ability, especially the understanding and generating ability, by fine-tuning these language-specific neurons.
domly sampled neurons may even increase the performance because irrelevant neurons are removed, which also aligns with the finding from Sharma et al. (2023). When assessing the differential impact on English and non-English language performance post-deactivation, specifically the difference calculated as \(\Delta_{\text{Eng}}-\Delta_{\text{n-Eng}}\), it is evident that the deactivation of random neurons within the understanding layer amplifies this effect. This observation lends partial support to the hypothesized role of the understanding layer in language processing.
Furthermore, researchers have found that "As LLMs progress at a rapid pace, existing datasets usually fail to match the models' ever-evolving capabilities, because the complexity level of existing benchmarks is usually static and fixed" [37]. To address this, work has been done to create dynamic benchmarks that are resistant to simple memorization [38; 37]. Researchers have also explored the idea of generating an entirely synthetic benchmark based on a user's specific environment or use case [14; 27]. While these techniques can help with contamination, decreasing the level of human involvement can pose additional risks regarding correctness and the ability to solve problems.
\(\bullet\)_Predictable scaling_. In practice, scaling law can be used to instruct the training of LLMs, and it has been proven feasible to reliably estimate the performance of larger models based on that of smaller models, called _predictable scaling_[46]. The benefits of predictable scaling for training LLMs are mainly twofold. Firstly, for large models, it is infeasible to rigorously examine various training tricks or variants, and it would be very helpful if experiences gained from small models could also apply to large models. For instance, small proxy models can be trained to find the optimal schedule of the data mixture for large models [59]. Secondly, the training of large-scale models takes a long time, often suffering from issues such as training loss spike, and scaling law can be employed to monitor the training status of LLMs, _e.g._, identifying abnormal performance at an early time. Despite that scaling law characterizes a smooth trend of performance increase (or loss decrease), it also indicates that _diminishing returns7_ might occur as model scaling. An empirical study [58] from the OpenAI team has shown that representation quality or semantic content can still effectively improve even if approaching the point of diminishing returns (_i.e._, approaching the irreducible loss) [58]. This finding suggests that training large models are promising for improving the performance of downstream tasks. To further explore scaling effect, a potential issue is that the amount of available data for training LLMs is actually limited. With the ever-increasing model scale, the public text data would be soon "exhausted" for LLMs [60]. Thus, it will be meaningful to study how scaling laws apply to a data-constrained regime [61], where data repetition or augmentation might be useful to alleviate data scarcity.
\(\bullet\)_Graph-structured reasoning._ Although the tree structure facilitates parallel reasoning, it also imposes restrictions on the reasoning process. With more complex topological structures, graphs offer greater flexibility in reasoning, enabling the characterization of more intricate relationships and interactions. For instance, Graph of Thoughts (GoT) [517, 518] conceptualizes the reasoning process as an arbitrary graph, where vertices denote intermediate thoughts and edges denote the interdependence between these thoughts. Compared with ToT, it can further utilize thoughts from other reasoning paths when generating new thoughts. However, such an approach requires a large number of interactions with LLMs, making the thought exploration process highly inefficient. To reduce potentially meaningless thought exploration, XoT [519] further proposes to guide the search of thoughts with pre-trained policy and value networks.
In a world torn by a devastating uav between two drive families, the Skywards and the Underlings, "Divine Rivals" follows the journey of Iris Winnow, an organized junior reporter at the Chuth Cagette in the city of O'A, as she strives to become a renowned journalist and uncover the truth about the uav that took her mother and brother, Forest, from her. The Skywards and Underlings' conflict has left the realm of Cambria in chaos, where magic and mythology intertwine with the mortal world, and gas walk among humans, Iris' dream of becoming a great journalist is fried by her love for her late brother and her desire to understand why he joined the uav. She faces face competition from her professional rival, Roman Kitts, at latented storyteller hailing from a wealthy family. While Iris longs to write about significant issues, Roman, was shares a complicated background with Iris, as asteroids in storytellig and is being presented by his father to join the family business. EdD Autry, the editor of the Oath Cagette, is reluctant to cover the urx, much to Iris's tisanya. Iris' family is burdened with financial difficulties and personal struggles after losing her mother and brother to the urx. Driven by her determination to uncover the truth about the uav and her little brother's involvement, Iris fuss so like in her resolute. The Individual Tribune, a rival newspaper, announces an opportunity for urx correspondents, enticing its complete with Roman for the position. Iris's journey is further complicated by her mysterious correspondence with an unknown person who lancse here letters in her wardede, forming a unique box. As Iris computes for the position, she reflects on her past and her friendship with Forest, who introduced her to the world of journalism. Their differing views on the urx and the Skyward goldes, Erwa, drove them apart. Iris wanted to highlight the theory of soldiers, public Forest aimed to reveal Enru's manipulation of young minds to let the urx. Despite the challenges, Iris remains focused on her goal and finally manages to land a job as a way to understand for the hridridation Tribune. She chunks on a dangerous journey to cover the urx, facing various charges and hardships along the way. Iris continues to write letters to the mysterious stranger, who becomes her confidante and support straining her perilous journey. Meanwhile, the town of Audon Bluff programs for an unending battle against an unknown enemy - Dacre, the god of the Underlings. Iris, Roman, and other companies gather in Audon Bluff to aid the town's defense, as tensions rise with theft of emergency supply kits and Keegan's distracted behavior. Amited the chaos, Iris, and Roman reliable her connection, and Roman propose to Iris, offering a moment of hope in the face of destination. As the group prepares for battle, Iris witnesses ffstand the horrors of war, including the loss of innocent lives and the destruction of her childhood home, Cliver Hill. Despite the pain, Iris continues to report on the urx, using her words to both document the battles and search for Forest. "Dicine Rivals" is a compelling narrative that explores themes of identity, courage, and the power of words in the face of war. Iris's unwarving determination to make a difference by writing about the truth, seeking answers about her brother, and making her urx has a journalist resonators throughout the story.
valuable as our system can be used to complement the crowd's prediction when there is greater uncertainty. When all three conditions are jointly met, our system beats the crowd significantly (by more than 1.5 standard errors in both Brier score and accuracy).
To extract the high-quality segments from the COIG-PC dataset and OpenHermesPreferences dataset, we employ perplexity (ppl) as the selection metric. Specifically, we use the Qwen-7B (Bai et al., 2023) model to compute the ppl for samples drawn from the SFT dataset. In our data filtering process for the SFT dataset, we retain only those entries with a perplexity score below 3,000 under Qwen-7B.
**Reasoning prompt.** We hand-craft 15 prompts for eliciting forecasts and reasonings. The best prompt (Figure 15) achieves a Brier score of 0.167 on the validation set (while fixing the optimal hyperparameter choices found by the optimization stages above). Two other top prompts get 0.170 and 0.174. The best prompt is given in Figure 15.
Question answering in conversational way naturally improves user experiences by addressing followup questions. The model can also raise clarification questions for users if necessary, which can reduce hallucination. Thus, it becomes the default format of deploying QA models in production (e.g. OpenAI, 2022; Google, 2023; Anthropic, 2023).
The multimodal RAG domain has experienced significant growth, highlighting a pivotal advancement at the confluence of text and visual comprehension. The introduction of MuRAGChen et al. (2022) marked a breakthrough by amalgamating textual and visual information for language generation, establishing a new standard for multimodal datasets. This model showcased the efficacy of utilizing a multimodal memory system to boost the accuracy in question-answering and reasoning tasks.
Large "instruction-tuned" language models (i.e., finetuned to respond to instructions) have demonstrated a remarkable ability to generalize zero-shot to new tasks. Nevertheless, they depend heavily on human-written instruction data that is often limited in quantity, diversity, and creativity, therefore hindering the generality of the tuned model. We introduce Self-Instruct, a framework for improving the instruction-following capabilities of pretrained language models by bootstrapping off their own generations. Our pipeline generates instructions, input, and output samples from a language model, then filters invalid or similar ones before using them to finetune the original model. Applying our method to the vanilla GPT3, we demonstrate a 33% absolute improvement over the original model on Super-NaturalInstructions, on par with the performance of InstructGPT\({}_{001}\),1 which was trained with private user data and human annotations. For further evaluation, we curate a set of expert-written instructions for novel tasks, and show through human evaluation that tuning GPT3 with Self-Instruct outperforms using existing public instruction datasets by a large margin, leaving only a 5% absolute gap behind InstructGPT\({}_{001}\).
Footnote 4: Another interpretation of privacy focuses on the sensitivity of the data involved, when a model is trained on and able to reproduce personal identifiers or other forms of “private data.” Our definition is more expansive.
better outcomes when both approaches are applied concurrently. One primary factor of this comparable performance may be that RFT does not sample extensively, yet the critical steps in the reasoning paths remain largely unchanged from the ground truth. The primary distinctions of sampled new paths lie in relatively minor semantic variances, without significantly increasing the new observed connections among local variables (Prystawski et al., 2023). It allows us to attain comparable effects by employing masked reasoning paths, where we sample new skip-connection by dropping the intermediate steps. On the other hand, if we compare from the perspective of addressing exposure bias, RL's self-sampling provides careful supervision for the error-prone inference distribution, whereas MFT aims to directly prevent the model from learning the bias present in the ground truth. We provide additional introduction about RL and RFT in Appendix B.
Secondly, with increasing vocabulary size, fertility and parity reduce in all cases, which can be explained by the tokenizer requiring fewer subword tokens when tokenizing text given a larger vocabulary. However, it can be observed that for monolingual English tokenizers, the fertility is less dependent on the vocabulary when tokenizing English documents, implying that 33k might be a sufficiently large vocabulary.
and makes a forecast; and (3) aggregation, which ensembles individual forecasts into an aggregated prediction. Each step makes use of an LM or a collection of LMs (either prompted or fine-tuned) (Figure 1).
Related Work.Theoretically, scaling laws have been derived in various settings: for non-parametric models (Schmidt-Hieber, 2017; Suzuki, 2019; Bordelon et al., 2020), in the kernel regime under the Gaussian design (Spigler et al., 2020; Cui et al., 2021, 2022; Maloney et al., 2022), or in memorization-like settings with discrete data (Hutter, 2021; Debowski, 2023; Michaud et al., 2023). Taking finite model capacity and optimization into account, Cabannes et al. (2023) recently proved scaling laws in constraint-capacity associative memories, and our Triplet Scaling Law builds on this work.
that uses RMSNorm but also biases (e.g. on the output logits). In Figure 5c we show that indeed the \(h\)th singular value does not decrease for this model that uses RMSNorm.
The best zero-shot and scratchpad prompts are shown in Figure 5 and Figure 6. In both prompting styles, models are only provided with the question, background, resolution criterion, and question's open and close dates (date_begin and date_end). All the data are sourced from the forecasting platforms and publicly available on the question page to human forecasters. We do no additional news retrieval.
**GPT-4 Judgement:** The first step is incorrect; '2/.5' suggests division when it should be multiplication, '2*.5'. The correct calculation for the reading time is 2/2 hours, which equals 1 hour, not '2/.5' hours. However, the numerical result '1 hour' happens to be correct despite the incorrect operation shown. The remaining steps correctly add the two activities to find the total daily time spent, correctly multiply by the number of occurrences per week, and correctly multiply by the number of weeks to find the total hours spent over four weeks.
\(\bullet\)_Visual instruction data._ Extensive work [856, 831] has empirically found that both quantity and quality of visual instructions have an important impact on model performance of MLLMs. One basic way to construct visual instructions is to leverage the exceptional capability of LLMs to synthesize instructions based on text descriptions of images [851]. To further enhance the quality of instructions, one can construct fine-grained visual instructions with the help of human annotation [857, 833] or synthesize more complex data through carefully-designed prompts [835]. Despite the effectiveness of the above LLM-based approaches, one primary question emerges as to whether a LLM (_i.e._, text generation model without training on any images) possesses the ability to generate sufficiently good visual instructions solely based on verbalized visual information (_e.g._, captions and coordinates). Specially, existing work has also revealed that visual instructions generated by LLMs sometimes contain misinterpretations about the visual information, _e.g._, object hallucination [844]. Therefore, it is crucial to design effective verification methods to control the quality of instruction data generated by LLMs [835]. Furthermore, it still needs more investigation about what makes good visual instructions and how visual instructions elicit specific multimodal abilities in MLLMs.
The shift from text unimodal to vision-text multi-modal introduces richer information, meanwhile bringing more challenges. Some works have attempted to explore X-of-thought reasoning in multi-modal scenarios by fine-tuning multi-modal models to generate a high-quality chain of thoughts.
* _The pipeline optimizes performance effectively._ The model trained on the optimized mixture can achieve a performance comparable to the one trained on the default mixture with only 73% steps. It eventually attains a performance that requires 48% more steps if trained using the default mixture. This indicates the effectiveness of our pipeline in mixture optimization. * _Predictions provide references for the balance of model capabilities._ Prediction results in Fig.
'distances' from the ground-truth text. Perhaps that iterative refinement also may be more difficult in the space of language model probability outputs than text embeddings, due to a lack of convexity; it is also plausible that a different architecture or set of hyperparameters may be needed to train a more powerful inverter using iterative refinement.
**Theorem 2.1**.: _Consider long-tail real-world data with exponent \(\beta>1\), and let the cutoff for AI-generated data be \(k\).
Reasoning TaskDeactivation methods: (i) randomly sampled neurons in the attention structure of task-solving layer. (ii) randomly sampled neurons in the task-solving layer. (iii) randomly sampled neurons in the task-solving layer. (iii) randomly sampled neurons in all layers. (iv) language-specific neurons in the task-solving layer. (v) language-specific neurons in the understanding layer and generation layer. (vi) language-specific neurons not in the attention structure of task-solving layers.
We define the problem of inversion from language model outputs and analyze inversion approaches from an attack and defense perspective. We show that this attack vector can be used to elicit hidden prompts from LM systems, even when we do not have direct access to model output distributions.
FILCO (Wang et al., 2023c) employs the Wikipedia dataset from the KILT benchmark, referred to as the "Wizard of Wikipedia" (WoW), to generate subsequent dialogue. This process involves basing the output on a Wikipedia article, with the input comprising the history of multiple rounds of dialogue. RA-DIT (Lin et al., 2023b) also employs the WoW dataset in the fine-tuning phase. As a result of the command tuning operation, the model outperforms Llama (Touvron et al., 2023a) and Llama-REPLUG (Shi et al., 2023b) with the same parameters for dialogue generation in the zero-shot condition. The incorporation of Selfmem (Cheng et al., 2024) into the retrieval-augmented generator markedly enhances the generation of dialogue, as a consequence of its remarkable flexibility.
* With 100 exposures, the capacity ratio degrades only by 2x (Figure 8(g) vs. Figure 8(a)). * With 300 exposures, the capacity ratio matches that of the 100-exposure scaling law without junk (compare Figure 8(h) with Figure 8(a)).
In our dataset, questions can get resolved long before their official close date. This occurs for questions like "Will \(\langle event\rangle\) happen by \(\langle date\rangle\)", where resolving early indicates that the event did occur (see Table 1 for an example). It is tempting to choose retrieval dates with respect to the resolve date so that each question can receive the same number of retrieval dates, e.g. by retrieving at geometric intervals between the open and resolve date. However, this would leak information, since the retrieval date would now depend on the resolve date, which, as we explained, correlates with the resolution.
**Impact of Datasets on Learning Dependency.** We investigate the impact of different datasets on the effectiveness of the MFT technique, specifically, which types of data are best suited for MFT
To find the value of \(f(-2)+f(-1)+f(0)\), we need to evaluate the function \(f(x)\) at these three values of \(x\).
When plotting the residuals of this fitted scaling law and those of our estimate of the same scaling law, it becomes clear that the estimated scaling law fails to fit the data well.
Converting to a fraction, or "a tenth of ten," so $10 2 - bucks$ = $1$ $10 - bucks$
For the Plus version, we utilize a larger LoRA rank compared to the basic version. Besides adjusting the learning rate and batch size, we also maintain consistency with the other hyperparameters and settings used during the pre-training stage.
We would have been able to use differential calculus to obtain the answer by calculating the direction of change if CnstScore\((\tau_{\text{2H}},\tau_{\text{1H}})\) were directly dependent on EntRec\({}^{l}(e_{2},\tau_{\text{2H}})\). However, there exists no direct functional dependency between the two values. Instead, we leverage the shared reliance of both metrics on \(\mathbf{x}^{l}\) for computation where \(l\in[0,L-1)\),3 redefining them as EntRec\((\mathbf{x}^{l})\) and CnstScore\((\mathbf{x}^{l})\) relative to \(\mathbf{x}^{l}\). This reparameterization allows us to change the question to: if EntRec\((\mathbf{x}^{l})\) is increased by altering \(\mathbf{x}^{l}\), does CnstScore\((\mathbf{x}^{l})\) also increase?
Mr. Price, from heart for their unconventional methods and reputation per unit, Jurning and Arrays include Mrs. Price, with 1984.
Some of the popular frameworks which are useful for LLM training includes (note that some of them can be used beyond LLM training too):
Instruction fine-tuning is crucial for getting the best performance out of the pre-trained decoder models across different tasks. We use a mix of general and legal instructions to train the model to understand and follow instructions well, with a focus on legal expertise.
**Capacity Enhancement**.
**AGI-Eval**(Zhong et al.2023): AGIEval uses questions from standardized tests taken by millions of people, including college entrance exams, law school admission tests, and professional qualification exams. It has 19 tasks in both English and Chinese.
To find the consecutive integers between which \(\log_{2}10000\) lies, we need to find the powers of \(2\) that are immediately less than and greater than \(10000\).
**Tree-Instruct**(Zhao et al., 2023) improves instruction quality by prompting the LLM to implicitly complicate instruction through its semantic tree. Following the original paper, we use the subsampled Alpaca dataset as seed data. We set the number of tree nodes to 10 for best possible performance.
**LangChain**[231] is a framework for developing applications powered by language models.
Given these distinctions, for LLaMA models, we use the notation LLaMA-\(\ell\)-\(h\) for \(\ell\) layers, \(h\) heads, and \(64h\) hidden dimensions; we omit group-query attention as LLaMA recommends it only for its 70B model. For Mistral, denoted as Mistral-\(\ell\)-\(h\), we enable group-query attention with 4 groups if \(h=0\pmod{4}\), 1 group for odd \(h\), or 2 groups otherwise.
\(\bullet\)_Pre-LN._ Different from post-LN, pre-LN [268] is applied before each sub-layer, and an additional LN is placed before the final prediction. Compared with post-LN, the Transformers with pre-LN are more stable in training. However, it performs worse than the variants with post-LN [269]. Despite the decreasing performance, most LLMs still adopt pre-LN due to the training stability. However, one exception is that pre-LN has been found unstable in GLM when training models more than 100B parameters [93].
**Skypilot**[228] is a framework for running LLMs, AI, and batch jobs on any cloud, offering maximum cost savings, highest GPU availability, and managed execution.
\(\bullet\)**JAX**[188] is a Python library for high-performance machine learning algorithms developed by Google, allowing users to easily perform computations on arrays with hardware acceleration (_e.g._, GPU or TPU). It enables efficient computation on various devices and also supports several featured functions, such as automatic differentiation and just-in-time compilation.
We then train new models (with the same architecture) to predict GCD, from AI data (generated by the above model), and compare to training with correct data - from correct computation of the GCD. When trained on small number of examples (less than \(100\) million), models learning from AI data achieve better accuracy (Table 2). We believe this is due to the fact that AI data smoothes away all the hard case, therefore presenting the model with a cleaner signal in the initial stages.
**Output** <p>Reference [1]<p>[ [IsRET] =Relevant] Rory Tapner is a British businessman who has served as the Chief Executive Officer of Coutts, a private banking and wealth management company, since 2010.[ [IsRET] =Contradictory], **Reference [2]<p>[ [IsRET] =Relevant] Rory Tapner was born on 30 September 1959 in London, England.[ [IsRET] =Fully Supported]
demonstrating "emergence" of good BLEU performance with scale. This type of "emergence" has been massively confirmed in (Wei et al., 2022), where a working definition of "emerging" is "not present in smaller models, but appears in larger models". In this sense, (Wei et al., 2022) demonstrate empirically a large number of "skills" appearing with scale, like Multi-Task NLU, Modular arithmetic, word unscrambling and transliteration.
ComputeContinuous pretraining utilizes \(256\) MI250 AMD GPUs. For instruction fine-tuning, workload distribution occurs across 16 MI250. Evaluation procedures are seamlessly conducted on a single MI250.
and Rho-1-7B achieved a state-of-the-art 40.6% and 51.8% on MATH dataset, respectively. On some unseen tasks (_e.g.,_ TabMWP and GSM-Hard), Rho-1 also demonstrates a certain degree of generalizability, with an average few-shot accuracy improvement of 6.2% on the Rho-1-Math-1B and 2.7% on Rho-1-Math-7B.
This algorithm is mainly used for very well-known models such as BERT and Electra. At the beginning of training, the algorithm takes all the alphabet from the training data to make sure that nothing will be left as UNK or _unknown_ from the training dataset. This case happens when the model is given an input that can not be tokenized by the tokenizer. It mostly happens in cases where some characters are not tokenizable by it. Similar to BytePairEncoding, it tries to maximize the likelihood of putting all the tokens in vocabulary based on their frequency.
Large language models (LLMs) have demonstrated remarkable open-domain capabilities. Traditionally, LLMs tailored for a domain are trained from scratch to excel at handling domain-specific tasks. In this work, we explore an alternative strategy of continual pre-training as a means to develop domain-specific LLMs. We introduce _FinPythia-6.9B_, developed through domain-adaptive continual pre-training on the financial domain. Continual pre-trained FinPythia showcases consistent improvements on financial tasks over the original foundational model. We further explore simple but effective data selection strategies for continual pre-training. Our data selection strategies outperforms vanilla continual pre-training's performance with just 10% of corpus size and cost, without any degradation on open-domain standard tasks. Our work proposes an alternative solution to building domain-specific LLMs from scratch in a cost-effective manner.
**Sequence Tagging.** The sequence tagging tasks, _e.g._, named entity recognition (NER) [755] and part-of-speech (POS) tagging [756], are also fundamental tasks. Typically, such tasks require assigning each token in the input sequence a proper semantic category label, _e.g._, the classic B-I-O (_Beginning, Inside_ and _Outside_) tagging scheme for NER tasks. In the era of deep learning, early efforts [757, 758] mainly integrate the learned sequence representations (_e.g._, using CNN, LSTM, and BERT) into the classic conditional random field model (CRF), which performs the tagging task based on structural prediction. Recently, researchers have tested the performance of LLMs in sequence tagging tasks, but observed that LLMs still face challenges in solving them using in-context learning [753], especially for special categories with ambiguous or rare names, _e.g._, the "MISC" (_miscellaneous entity_) and "ORG" (_organization_) classes. A possible reason is that LLMs may misunderstand the meanings of these classes in the human-annotated dataset, making it difficult to accurately understand their semantics according to the instruction and limited examples in the context.
erroneous because another model is used to judge. Still, even today, evaluating purely generated content is very hard and no completely fitting metric is not found, metrics are either looking for simplistic features such as N-Gram, SkipGram, etc, or they are models with unknown accuracy and preciseness [204].
To further validate the behavior of Adapt-LLM when requesting additional context, Figure 2 illustrates the proportion of questions for which our model generates the \(\langle\text{RET}\rangle\) token, aggregated by popularity score intervals (left image for Adapt-LLM trained on NQ and right image for SQuAD). Mallen, Alex Troy and Asai, Akari and Zhong, Victor and Das, Rajarshi and Khashabi, Daniel and Hajishirzi, Hannaneh [22] suggest that high-popularity questions can be adequately answered using the parametric memory of the LLM, while lower popularity scores necessitate extra context. In Figure 2, we observe this pattern for both versions of Adapt-LLM, indicating that our model, despite lacking access to popularity scores during training or inference, has learned effective criteria for requesting additional context.
Comparison of Synthetic SFT Data versus Real DataTo assess the quality of the synthetically generated math questions, we evaluate their effectiveness against real questions from the GSM8K and MATH training sets, utilizing a LLaMA-2 7B model, as detailed in Table 1. The results indicate that the synthetic math questions are nearly as effective as the real ones.
What if the model is not sufficiently trained? For instance, there might be instances where knowledge appears only 100 times throughout the pretraining phase. We also calculate the capacity ratios for models trained with 100 exposures on \(\mathsf{bioS}(N)\).
Tokenizers in Decoder ModelsAn overview of current mono- and multilingual LLMs is provided in (Lin et al., 2022; Shliazhko et al., 2022; Scao et al., 2022). Stollenwerk (2023) evaluated the intrinsic metrics of the GPT-SW3 (Eggren et al., 2023) tokenizer that focused on the Nordic languages. As part of their work, Shliazhko et al. (2022) ablated different tokenizer pre-processing approaches while keeping the tokenizer algorithm, the vocabulary size, and the employed implementation fixed. In none of the other major LLM publications, the extrinsic tokenizer performance has been studied.
Since the aggregate of predictions is usually superior to individual forecasts (Tetlock and Gardner, 2015), we elicit multiple predictions from the base and fine-tuned models.
Figure 7: On the left, we plot the singular values that are extracted using our attack on GPT-2 small—the estimated hidden dimension is near 768. On the right, we post-hoc extend the dimensionality of the weight matrix to 1024, as described in Section 8. This misleads the adversary into thinking the model is wider than it actually is.
**What kinds of faithfulness errors do LLM summarizers make? (SS3)** A qualitative analysis of Tables reveals that the majority of claims marked as unfaithful are related to _events_ or _states_ of characters and relationships. Furthermore, most of these claims can only be invalidated via multi-hop reasoning over the evidence, highlighting the task's complexity and its difference from existing fact-verification settings (Min et al., 2023; Kamoi et al., 2023).
\(\bullet\)_Input data._ In common cases, it is straightforward to describe input data (_e.g._, an instance to be responded by LLMs) in natural language. For special input data, such as knowledge graph and table, it is necessary to apply an appropriate and convenient way to make them readable for LLMs. For structured data, linearization is commonly used to transform the original records (_e.g._, knowledge triples) into sequences [458] due to the simplicity. Further, the programming language (_e.g._, executable code) has also been utilized to formulate the structured data, which can also support using external tools (_e.g._, program executor) to produce the precise results [459, 460].
Large Language Models (LLMs) have demonstrated impressive performance across various tasks Brown et al. (2020). There is an increasing trend of releasing pre-trained LLMs and fine-tuned variants Touvron et al. (2023, 2023). Many of these fine-tuned variants aim to augment the knowledge or linguistic capabilities of the existing LLM Roziere et al. (2023); Cui et al. (2023).
CC-RecentNewsWe first construct CC-RecentNews, a novel text corpus containing relatively _new_ knowledge as \(D_{1}\). We use news-please (Hamborg et al., 2017), similar to the CC-NEWS (Liu et al., 2019) and REALNEWS dataset (Zellers et al., 2019), to crawl 221,779 news articles published from May 2020 to April 2021. LMs initially pretrained on \(D_{0}\) constructed before May 2020 can be continually pretrained on CC-RecentNews to gain relatively _recent_ world knowledge.
The passkey task hides a random number into a long text and asks it back at the model output. The length of the distraction text is varied by repeating a text chunk multiple times. The previous work (Chen et al., 2023a) showed that a 8B LLMA model can solve the task up to 32K length when fine-tuned with the same 32K length inputs with Position Interpolation. We take this challenge further and fine-tune on only 5K length inputs to test on 1M length regime.
1 to 3 points: The response lacks basic consideration of security, with little to no identification of potential risks and no effective preventative measures provided.
A more surprising finding is that the performance of some of the parameter-expansion methods are even higher than that of T5-Initial, which is considered to be the upper bound for KILT because T5-Initial is only trained on behalf of the knowledge from \(D_{0}\). For example, T5-Modular shows higher scores than T5-Initial on 6 out of 11 tasks. Since the parameter-expansion methods force the model to store the new knowledge in the newly added parameters during continual pretraining, one careful conjecture is these LMs have learned to combine and utilize in its internal representation of both old and new knowledge stored in separate parameters during finetuning to maximize the performance.
Secondly, as an orthogonal direction, our method did not explore robustness of the instruction-tuned model towards adversarial attacks such as prompt injection (Liu et al., 2023) and jailbreaking Zou et al. (2023). In practice, we should apply adversarial defense techniques (Jain et al., 2023) accordingly to the instruction-tuned LLM from our method.
The State Space Model originates from Kalman filtering [35] which mainly introduces a linear filtering and prediction method. Kalman filtering can be divided into two steps, i.e., the prediction and correction step. The prediction is to estimate the current state based on the state of the previous time, and the correction is to estimate the optimal state by integrating the estimated state and the observed state of the current time. The State Space Model is a mathematical model that describes the behavior of a dynamic system using a set of first-order differential equations (continuous-time systems) or difference equations (discrete-time systems) to represent the evolution of the internal state of the system, and another set of equations to describe the relationship between the state and the output of the system. These equations can be expressed in matrix and vector form to deal with multivariable systems. Subsequently, Gu et al. [34] introduces a Linear State Space Layer (LSSL) that combines the advantages of recurrent neural networks (RNNs), temporal convolutional networks, and neural differential equations (NDEs) while addressing their shortcomings in model power and computational efficiency. This new sequence model is inspired by control systems and implemented through the linear state space layer (LSSL).
Figure 19: A response from our fine-tuned model on the question “Will Israel carry out and explicitly acknowledge a deadly attack on Iran before 2024?” (resolved NO). In our fine-tuning phase, one of the prompts we use to generate strong reasonings asks the model to build a decision tree and assign probabilities. The fine-tuned model learns the reasoning path (without being explicitly prompted to do so).
In terms of visual and linguistic information fusion, Zhao et al. [147] optimizes the internal information integration of the Mamba language model to achieve a more effective expression. Decision Mamba(DMamba) [148] integrates the Mamba framework into the Decision Transformer (DT) [152]. A series of experiments comparing Decision Mamba and DT show that Mamba is effective in reinforcement learning (RL) tasks. But simply applying Mamba blocks to DT does not improve efficiency, because the RL tasks considered by the authors have a large number of CPU and GPU interactions. Another deficiency is the absence of a hyper-parameter search and an analysis of how to use the Mamba block more effectively to reflect the data structure of RL tasks. Sigma [149] is the first State Space Model successfully applied in multi-modal semantic segmentation. It is composed of VMamba, an attention-based Mamba fusion mechanism and a channel-aware Mamba decoder, and has shown excellent performance in various experiments. However, Sigma is underutilized in handling longer sequences, and the memory consumption of Mamba encoders is still relatively large, making it difficult to deploy on lightweight edge devices.
\(\bullet\)_LLaMA_[57] extracts training data from various sources, including CommonCrawl, C4 [82], Github, Wikipedia, books, ArXiv, and StackExchange. The training data size for LLaMA (6B) and LLaMA (13B) is 1.0T tokens, while 1.4T tokens are used for LLaMA (32B) and LLaMA (65B).
For each pair of documents identified as a potential match, more computationally expensive similarity metrics can be employed as a subsequent filtering step.