Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
Footnote 4: They consider two types of training, pre-train + finetune (FT) in which they pre-train the model with biography entries and then finetune with QA tasks; mixed-training in which they train the model once with both the biography entries and the QA tasks (not in the same context). They always use half of the individuals’ QA tasks for training and evaluate the QA accuracies on the remaining half.
only degrades the overall logical reasoning performance, but also escalates the premise order effect. The _Reversal Curse_(Berglund et al., 2023) unveils another perspective of the order effect, where they show that an LLM that recognizes "A is B" does not necessarily learn that "B is A." While their work studies the order effect between two entities within a single factual statement, our work focuses on reasoning problems with multiple premises, without restrictions on the number of (or relationship between) entities. In particular, for logical reasoning, we demonstrate that random permutations of premises often result in **worse** accuracy than the purely backward order.
The OSCE is typically divided into multiple stations (often 8-12), each simulating a real-life clinical scenario enacted by standardized patient actors trained to portray specific symptoms or conditions based on pre-defined scenario descriptions. At each station, students are given specific tasks to perform, such as taking a clinical history, or making a diagnosis. Each station has a set time limit, ensuring fairness and efficient assessment. Trained examiners observe students' performance at each station using a pre-defined checklist or marking scheme. They assess clinical skills like communication, history-taking, physical examination techniques, clinical reasoning, and decision-making.
[37] use Brier Score [2] as the metric for BIG-Bench. We argue that increase in Brier Score does not always represent improvement of performance on the multi-choice task, since Brier Score is also related to the allocation of probabilities for incorrect options. For example, questions in the MMLU dataset have four options (A, B, C, and D) and the frequency of the four options as correct is equal. Consider two models that give the same probability independent of questions. One model predicts \((1,0,0,0)\) for the four options and the other model predicts \((0.25,0.25,0.25,0.25)\). The Brier Score for the former is 1.5 while the Brier Score for the latter is 0.75. However, both models do not learn the relationship between questions and correct options at all. One can argue that the latter model better fits the distribution of correct options in the dataset, but the improvement is not as large as the different of 1.5 and 0.75.
Figure 5: SFT data scaling curve. Compared with UltraChat and its cleaned version UltraChat 200K, our SFT data demonstrates clear scaling advantages. We attribute its steep slope to the data quality.
We consider three different LM sampling mechanisms as defenses against prompt inversion: adjusting the softmax temperature during sampling, adjusting the top-p parameter of nucleus sampling (Holtzman et al., 2020), and adjusting the total number of tokens considered (top-K). We sample from LLama-2 7b (non-chat) and feed probabilities into our inverter model according to the desired strategy. Results are visualized in Figure 3.
In contrast to the latest LLM-based generalist solution (e.g., OpenAI, 2022), most of previous studies focus on fine-tuned expert models on specific domains or datasets (Feng et al., 2020; Izacard & Grave, 2021; Chen et al., 2022; Gao et al., 2022; Nakamura et al., 2022; Adlakha et al., 2022; Wu et al., 2023).
For language identification analysis, we utilize the FastText Joulin et al. (2016, 2016) language identification model to detect the language of the generated tokens. As for repetition analysis, we assess the proportion of duplicated n-gram tokens at the BPE level within the combination of the generated output and the prompt.
A key aspect of Reflection is the LLM's capacity for self-editing. By evaluating its initial response, the model can identify potential errors or areas of improvement. This iterative process of generation, reflection, and revision enables the LLM to refine its output, enhancing the overall quality and reliability of its responses.
The evaluation of LLMs poses particular challenges due to the evolving landscape of their applications. The original intent behind developing LLMs was to boost the performance of NLP tasks such as translation, summarization, question-answering, and so on [178]. However, it is evident today that these models are finding utility across diverse domains including code generation and finance. Moreover, the evaluation of LLMs encompasses several critical considerations such as fairness and bias, fact-checking, and reasoning. In this section, we outline the commonly used benchmarks for assessing LLMs. These benchmarks are categorized based on training or evaluating the LLM Capabilities.
Figure A16: Comparison data selection methods when repeating data at the 125M, 3B token budget scale. The x-axis is data selection ratio \(R\), and the y-axis is average perplexity on validation sets. We observe that selecting data to repeat via D4 outperforms other data selection methods, especially at low selection ratios \(R\) (note that low selection ratios in the fixed-data regime correspond to more epochs).
6. We have another instance of \(f\) and \(g\). Since \(f\) is odd and \(g\) is even, their product is an odd function. So, we can rewrite the expression as \(-[f(g(f(x)))]\).
After creating the dataset, we utilized GPT-4 to generate augmentations of the dataset. We instructed GPT-4 to provide paraphrased versions of the input data that fully retain the information while being reworded. Each paraphrasing iteration was done with a different seed to ensure variety.
- She has 4 less than a third as many cherry trees as apple trees. A third of the apple trees is 80/3, which is approximately 26.67.
This work introduces an efficient method to scale Transformer-based Large Language Models (LLMs) to infinitely long inputs with bounded memory and computation. A key component in our proposed approach is a new attention technique dubbed Infini-attention. The Infini-attention incorporates a compressive memory into the vanilla attention mechanism and builds in both masked local attention and long-term linear attention mechanisms in a single Transformer block. We demonstrate the effectiveness of our approach on long-context language modeling benchmarks, 1M sequence length passively context block retrieval and 500K length book summarization tasks with 1B and 8B LLMs. Our approach introduces minimal bounded memory parameters and enables fast streaming inference for LLMs.
When compared to other tree-based methods, LATS implements a self-reflection reasoning step that dramatically improves performance. When an action is taken, both environmental feedback as well as feedback from a language model is used to determine if there are any errors in reasoning and propose alternatives. This ability to self-reflect combined with a powerful search algorithm makes LATS perform extremely well on various tasks.
However, similar to the RNN model, we face the dilemma that the computation cannot be _parallelized_.
**Validation Perplexity.** In Figure 2, we show that a relatively small amount of data selection using any of the three methods (small \(R\)) brings consistent improvements on all validation sets. However, as we increase \(R\), we observe _opposing effects_ on web snapshot and non-web-snapshots validation sets. We analyze this discrepancy in-depth in Section 4.4. However, on the Instruct OPT validation set, which corresponds much more closely to the the high-quality generations we want our LLMs to achieve, we found that all three methods led to consistent and clear perplexity improvements. Notably, we found that while all three methods provided benefits, D4 outperformed using both SemDeDup and SSL Prototypes independently, with the most notable gains exhibited when the source dataset is around 4x the target dataset size. Given that D4 consistently improves with source dataset size, we estimate this gap to grow with source dataset size.
* **IE-Based Metrics**: Utilize Information Extraction models to simplify knowledge into relational tuples, then compare these with the source. * **QA-Based Metrics**: Assess the overlap between generated content and the source through a question-answering framework (see [152]). * **NLI-Based Metrics**: Use Natural Language Inference datasets to evaluate the truthfulness of a generated hypothesis based on a given premise (see [153]). * **Faithfulness Classification Metrics**: Offer a refined assessment by creating task-specific datasets for a nuanced evaluation (see [154]).
Task-adaptive pre-training.Continual pre-training of language models on unlabeled data for a given task has been demonstrated to be beneficial for enhancing end-task performance [1; 8; 13]. In scenarios involving domain shift, domain-adaptive pre-training bears similarities to task-adaptive pre-training to some extent. Aharoni et al [1] documented that continual pre-training a model on a similar domain contributes to improved task performance on the target domain. Notably, the work closest to ours is presented in [8], which shows that continual pre-training of language models on both unlabeled task data and augmented unlabeled task data, sampled from the in-domain corpus based on similarity. While these works use task data, we also propose a task agnostic method, ETA-DACP, as task similarity is prohibitively expensive for LLMs.
Contributors to **model training and architecture** include Shane Arora, Iz Beltagy, Akshita Bhagia, Matthew E. Peters, Dirk Groeneveld, Ananya Harsh Jha, William Merrill, Jacob Morrison, Niklas Muennighoff, Dustin Schwenk, Saurabh Shah, Pete Walsh, and Mitchell Wortsman.
with \(\alpha\) controlling the steepness of the inverse square root decay. We shift and stretch the Inverse Square root decay to adapt to the interval \((t_{cd},t_{const}]\).
When we aim to enhance the capabilities of a LLM in a specific language, vocabulary extension is an intuitively reasonable approach. In this section, we evaluate the impact of vocabulary extension through the LLM-Eval benchmark, and the experimental results are presented in table 1. Initially, we collected one million Chinese sentences from the internet (approximately 0.5 billion tokens) and further pretrain the original LLaMA without vocabulary extension. Surprisingly, we find that this model significantly outperform the vocabulary-extended Chinese LLaMA, across settings of 1K, 5K, and 950K instruction tuning. This discovery is thought-privoking, given that the Chinese LLaMA underwent further Chinese pretraining on 30 billion tokens, a much larger volume than our 0.5 billion tokens. Moreover, within the 950K setting, we include results from extending the vocabulary on original LLaMA and training it with the same 0.5 billion tokens, to mitigate the influence of training data discrepancy. The outcomes remain consistent. This indicates that vocabulary extension is not a favorable choice within training scales of tens of billions of tokens. While we don't negate the efficacy of vocabulary extension in settings involving larger-scale pretraining (such as trillions of tokens), as reported in other literatures [23], this already leans more towards retraining than mere language transfer.
Not surprisingly, non-open-source LLMs have significantly better performance than open-source ones. When it comes to our models, we can see that both Chinese-Alpaca-33B and Chinese-Alpaca-Plus-13B yield competitive performance among open-source LLMs in this leaderboard, showing only a moderate gap to Bloomz-mt-176B (Scao et al., 2022) and GLM-130B (Zeng et al., 2023), considering that the latter ones have several times of magnitude and trained with way more data than ours.
* Figure 10 visualizes the location mentions in all the questions from our full dataset. * Table 11 gives the distribution of questions and forecasts across platforms in our full dataset. * Table 12 showcases a complete data sample in our curated set. * Table 13 shows a list of questions with how community predictions shift over time. * Figure 10(a) shows the opening dates of the questions in the full dataset. * Figure 10(b) shows the percentage of questions that receives the retrieval date at index \(k=1,2,3,4,5\).
distances between keys and queries. Different from the relative positional embedding methods like T5 [82], the penalty scores in ALBi are pre-defined without any trainable parameters. Empirical results in [264] have shown that ALBi has a better extrapolation performance on sequences that are longer than those for training than several popular position embedding methods such as sinusoidal PE [22], RoPE [263], and T5 bias [82]. In addition, it has been shown that ALBi can also improve training stability in BLOOM [78].
1. **Intrinsic Hallucinations**: These directly conflict with the source material, introducing factual inaccuracies or logical inconsistencies. 2. **Extrinsic Hallucinations**: These, while not contradicting, are unverifiable against the source, encompassing speculative or unconfirmable elements.
_Remark A.2_ (parameters).: These optimization parameters are _very natural_, as it is generally impossible to have a fixed set of parameters for model sizes across a large multiplicative range.
Proof of Theorem 3.2.: Let us first construct the domains \(\mathcal{Q}_{i}\)'s in Lemma F.1.
Footnote 5: Note that according to our third assumption, the quality of the training data is uniform and independent of the size.
Figure 1: Comparative performance analysis of Baichuan-7B, DeepSeek-7B, and DeepSeek-67B models across MMLU, CMMLU, and CEval benchmarks.
**Mixture-of-Experts.** For the above three types of architectures, we can further extend them via the mixture-of-experts (MoE) scaling, in which a subset of neural network weights for each input are sparsely activated, _e.g.,_ Switch Transformer [25] and GLaM [112]. The major merit is that MoE is a flexible way to scale up the model parameter while maintaining a constant computational cost [25]. It has been shown that substantial performance improvement can be observed by increasing either the number of experts or the total parameter size [246]. Despite the merits, training large MoE models may suffer from instability issues due to the complex, hard-switching nature of the routing operation. To enhance the training stability of MoE-based language models, techniques such as selectively using high-precision tensors in the routing module or initializing the model with a smaller range have been introduced [25]. More recently, there is widespread speculation that GPT-4 has been developed based on the MoE architecture, but without official verification.
**Performance on different skills.** Figure 6 compares _WizardLM_ and ChatGPT's skill levels on _Evol-Instruct_ testset. The result indicates that _WizardLM_ achieves 78% of ChatGPT's performance on average, with almost more than 90% capacity on 17 skills. However, _WizardLM_ struggles with code, math, and reasoning scenarios, revealing a noticeable gap with ChatGPT.
Experimental ResultWe first focus on the performance on zero-shot Relation Extraction (zsRE), which is measured on the dev set of 12 relations that are ensured to have no overlap with the 84 relations of the train set (Levy et al., 2017). Since the setting is similar to the zero-shot probing setting of IL, the trend of the result on the two datasets are similar. The performance of T5-Vanilla drops to half from that of T5-Initial as shown in IL, and the best performing method for both datasets is T5-Modular. In addition, corresponding with results from the CKL benchmark, parameter-expansion methods generally show stronger performance than the other methods.
Knowledge distillation is the process of learning from a larger model [143]. Earlier days of best-performing models release have proven that this approach is very useful even if it is used in an API distillation approach. It is also referred to as an approach to distill the knowledge of not a single model but in fact multiple models into a smaller one. Creating smaller models by this approach yields smaller model sizes that can be used even on edge devices. Knowledge distillation as shown in Fig 35, illustrates a general setup of this training scheme.
The results indicate that our simple sampling-and-voting method for instantiating agents can generally improve the performance of LLMs by increasing the ensemble size. Importantly, this method is orthogonal to different existing methods, which can lead to further improvements when combined with them.
**How many instances are there in total?** There are 43,090 instances in total. There are between 1,810 and 27,250 instances per hazard category.
Table 4 shows the list of 28 time-invariant relations of InvariantLAMA. We manually filter the 44 original LAMA relations to leave only the time-invariant relations. Templates such as "[X] works for [Y]." and "[X] is a member of [Y]." are excluded because the answer may change for different timestamps. In the template, [X] and [Y] refers to subject and object labels, respectively. Given a template with only the subject included, the model has to predict the object label [Y] for knowledge probing.
Figure 6 indicates a clear relationship between the size of the training data and the performance scores across the CEval, CMMLU, and MMLU benchmarks. All three subfigures exhibit a monotonically increasing trend, where the scores improve as the number of training checkpoints increases. This suggests that larger training datasets enable the models to learn more effectively, leading to better performance on evaluation tasks. The rate of score increase appears to be diminishing as the number of checkpoints grows. Initially, there is a steep improvement in scores, which gradually levels off. This could indicate that while larger training datasets initially provide substantial new information for the models to learn from, the incremental benefit decreases as the model approaches an asymptotic performance ceiling.
Figure 7: Xwin-Math’s aggregate performance on these two benchmarks is second only to GPT-4, demonstrating our model’s robust generalization capabilities.
SURGE (Kang et al., 2023) achieves customization through the application of graph-text contrastive learning. This method ensures that the generated dialogue responses are in tight alignment with the knowledge contained in the retrieved subgraph, yielding responses that are specific, relevant, and deeply rooted in the dialogue context. By maintaining consistency between the retrieved knowledge and the generated text, SURGE is capable of producing outputs that precisely reflect the detailed knowledge of the subgraph, enhancing the relevance and specificity of the responses.
Unlike pre-training, instruction tuning is often more efficient since only a moderate number of instances are used for training. Since instruction tuning can be considered as a supervised training process, its optimization is different from pre-training in several aspects [69], such as the training objective (_i.e.,_ sequence-to-sequence loss) and optimization configuration (_e.g.,_ smaller batch size and learning rate), which require special attention in practice.
In the following sections, we describe the real-world datasets, simulated self-play environment, fine-tuning process, and inference time chain-of-reasoning that we designed to optimize AMIE for diagnostic conversation capabilities and clinical communication skills.
We treated the last numerical value appearing in the output as the answer. We needed to adopt this heuristic because we are merging multiple models that were trained in different formats, which made it difficult to correct the output format. This method appeared to extract the answers correctly in almost all cases. Additionally, to determine the language of the output, we utilized fasttext [23; 24]. We used greedy sampling for generation and calculated the zero-shot pass@1 accuracy.
**Improvement Strategies.** Although real-world instructions from human users are more suitable for fine-tuning LLMs, it is difficult to collect them at a large scale. As alternatives to human-generated instructions, most existing research mainly adopts synthetic instructions generated by LLMs. However, there are some potential problems with synthetic instructions, such as poor topic diversity and uneven instruction difficulty (either too simple or too difficult). Thus, it is necessary to improve the quality of the synthetic instructions.
Taking \(T=T_{real}+T_{AI}\) and \(\pi=T_{real}/T\), we have the following important corollary of Theorem 3.2.
More examples on ChatQA-13B, ChatQA-70B, GPT-3.5-turbo, and GPT-4 can be found in Table 10.
Notice that when \(K=2\), Equation 18 reduces to the Bradley-Terry model. However, for the general Plackett-Luce model, we can still utilize the results of Eq. 5 and substitute the reward function parameterized by its optimal policy.
**Training overview.** Self-Rag enables an arbitrary LM to generate text with reflection tokens by unifying them as next token predictions from the expanded model vocabulary (i.e., the original vocabulary plus reflection tokens). Specifically, we train the generator model \(\mathcal{M}\) on a curated corpus with interleaving passages retrieved by a _retriever_\(\mathcal{R}\) and reflection tokens predicted by a _critic_ model \(\mathcal{C}\) (summarized in Appendix Algorithm 2). We train \(\mathcal{C}\) to generate reflection tokens for evaluating retrieved passages and the quality of a given task output (Section 3.2.1). Using the critic model, we update the training corpus by inserting reflection tokens into task outputs offline. Subsequently, we train the final generator model (\(\mathcal{M}\)) using the conventional LM objective (Section 3.2.2) to enable \(\mathcal{M}\) to generate reflection tokens by itself without relying on the critic at inference time.
Despite our diligent RAI efforts, as with most LLMs, there remains challenges around factual inaccuracies (or hallucinations), reproduction or amplification of biases, inappropriate content generation, and safety issues. The use of carefully curated training data, and targeted post-training, and improvements from red-teaming insights significantly mitigates these issues across all dimensions. However, there is significant work ahead to fully address these challenges.
**Monotonic Improvement** This approach resulted in notable improvements in our results, showcasing a direct correlation between the number of paraphrases utilized and the models' accuracy. Our experimentation revealed a compelling trend, shown in Figure 4. For all models tested, the accuracy was a monotonically increasing function of the number of paraphrases used. This observation strongly suggests the positive impact of paraphrase augmentation, yielding information repetition, on the model's ability to comprehend and generalize new knowledge from limited data.
Figure 1: **Overview of Evolutionary Model Merge**. Our approach encompasses (1) evolving the weights for mixing parameters at each layer in the parameter space (PS); (2) evolving layer permutations in the data flow space (DFS); and (3) an integrated strategy that combines both methods for merging in both PS and DFS. Notice that merging in the PS is not simple copying and stitching of the layers parameters, but also mixes the weights. This merging is akin to blending colors as illustrated here (e.g., red and blue becomes purple). Note that we translated the questions to English for the reader; the models operate on Japanese text.
**Early Explorations**. According to one interview with Ilya Sutskever14 (a co-founder and chief scientist of OpenAI), the idea of approaching intelligent systems with language models was already explored in the early days of OpenAI, while it was attempted with recurrent neural networks (RNN) [121]. With the advent of Transformer, OpenAI developed two initial GPT models, namely GPT-1 [122] and GPT-2 [26], which can be considered as the foundation to more powerful models subsequently _i.e._, GPT-3 and GPT-4.
since \(k^{-(\beta-1)}\lesssim T^{-(\beta-1)/\beta}=T^{-c}\). Putting things together then gives the claimed result.
GPT-4-Turbo _In the epic narrative of "Divine Rivals,"_ _are are introduced to Iris Winnow, a tenacious young woman propelled into the role of a war correspondent amidst a devastating conflict influenced by the ancient deities Etru and Dacre. Iris's journey into the heart of battle is not just professional but deeply personal, fueled by the disappearance of her beloved brother, Forest, who had enlisted to fight, and the anguish over her mother's demise. Her quest for truth and understanding is set against the backdrop of the war-torn Avalon Bluff, close to the war's front lines, where she encounters the stark realities of conflict, from the silent hope of a garden amidst destruction to the horror of wounded soldiers and the terror of artillery fire. Roman Kitt, previously Iris's rival at the Cath Gazette, emerges as a significant figure in her narrative, joining her as a fellow correspondent. Despite their history of competitive antagonism, their relationship evolves, highlighted by Roman's mysterious decision to leave behind a promising career and a broken engagement to be by Iris's side. Their dynamic is further complicated and enriched by their shared experiences on the battlefield and Roman's unspoken feelings for Iris, culminating in a poignant moment of sacrifice during a trench attack where Roman shields Iris with his body from a grande explosion. Adding a magical element to the story, Iris maintains a correspondence with Carver, a mysterious figure, through enchanted typewriters. These letters offer Iris solace and a metaphysical connection that transcends the chaos surrounding her, underscoring themes of hope and connection. As the narrative unfolds, shocking revelations come to light. Forest, Iris's brother, believed to have been fighting for Enro, is discovered to have been forcefully converted by Dacre after a near-death experience. This discovery updfalls through a dramatic reunion where Forest saves Iris from one of Dacre's deally gas attacks, only to reveal his changed allegance by decaying her and leaving Roman behind, presumed dead or captured. The heart of the story explores Iris and Roman's profound relationship, against the backdrop of war and dicine manipulation. Moments from their past, including their secret wedding in a garden before a devastating attack by Dacre's forces, and Roman's heartfelt letter confessing his love, serve as poignant reminders of what they fight for. This emotional core is juxtaposed with the chaos of the present, wherein Iris must confront the bertual of her brother and the uncertainty of Roman's fate. "Divine Rivals" delves into themes of love, loyalty, and the harrowing choices faced by individuals caught in the machinations of gods and war. Iris, within the turmoil of her personal and professional world, remains determined to uncover the truth, seeking to reclaim agency and hope amidst depsur. Her journey is not just a search for her brother and Roman but an exploration of the resilience of the human spirit in the face of dicine conflicts and the devastation of war.
Footnote 17: $64.6 for Claude-3-Opus, $169.4 for GPT-4, $47.5 for GPT-4-Turbo, $2.8 for GPT-3.5-Turbo, and $3.4 for Mixtral. All in USD.
\(\bullet\)_Simply increasing the number of instructions may not be that useful, and balancing the difficulty is not always helpful._ As the results shown in Table IX, balancing the difficulty and increasing the number of fine-tuning instructions are not very helpful in our experiments. Especially for scaling the instruction number, it even hurts the performance, _e.g._, a decrease from 29.81 to 26.63 in BBH for LLaMA (7B). It shows that simply scaling the number of synthesized instructions without quality control may not be effective to improve the performance. Furthermore, fine-tuning with the instructions of moderate difficulty also performs well in the chat setting, while slightly decreasing the performance in the QA setting. A possible reason is that we filter complex and hard instructions with large perplexity scores, hurting the model performance in answering complex questions.
_Remark_ G.1.: Due to the significant overlap among textbooks, especially those designed for PreK-12 education, estimating the total amount of knowledge contained within all English-language textbooks can be challenging. However, we attempt to do so as follows.
However, when we fit based on Equation. 3 to the first four points, as shown in dashed lines, we found that the latter three points were significantly below the curve. We believe this phenomenon may be related to the smaller proportion of more complex problems in the training data. Therefore, we resampled the 960K synthetic data according to the number of sentences in CoT solution. As can be seen from Figure 4 (right), when the proportion of complex problems is increased, the accuracy for simpler problems remains virtually unchanged, but the accuracy for more complex problems can be significantly improved. Moreover, the utilization of data resampling can increase the model's PassRatio@256 from 71.1 to 72.8. This experimental result provides new insights into data selection for mathematical reasoning tasks.
Based on the origin and variations of SSM mentioned above, many SSM-based works are constantly emerging, including but not limited to natural language processing, computer vision, and so on. The following summaries will respectively introduce the expansion and application of various fields.
It is not straightforward that the loss of LMs decides the performance on downstream tasks. For simplicity, we consider the Exact Match (EM) metric with single-token target. The score \(\mathrm{EM}(\hat{y},y)\) for the prediction \(\hat{y}\) of the prompt \(x\) given the ground truth \(y\) is 1 if \(\hat{y}=y\) and 0 otherwise.
1. Violent crimes 2. Non-violent crimes 3. Sex-related crimes 4. Child sexual exploitation 5. Indiscriminate weapons, Chemical, Biological, Radiological, Nuclear, and high yield Explosives (CBRNE) 6. Suicide & self-harm 7.
The size and complexity of a model are key factors affecting its learning capabilities and performance in reasoning tasks, yet the application of specific techniques may allow smaller-scale models to match or even surpass the capabilities of larger models. As shown in Figure 4 and Figure 5, by examining the performance on MATH, GSM8K, PIQA, HellaSwag, and WinoGrande across models of different sizes, it is observable that all models improve their abilities in tasks involving math, physical interaction understanding and commonsense reasoning in a relatively synchronized manner. Generally, the performance of 67b models surpasses that of 34b, which in turn exceeds that of 7b models at the same training phase in math-related datasets. However, exceptions like the OpenLLaMA-7b model approaching or even exceeding the capabilities of Yi-34b demonstrate that the model scale should be chosen based on the complexity of the task and the required reasoning capabilities.
Cluster-based FiltersWe further use unsupervised semantic clustering to group web documents. This clustering process enables efficient identification and analysis of documents sharing similar semantic features. The clustered data are subsequently annotated with quality labels, providing essential references for the optimization of Yi's data mixture strategy. Documents identified as low-quality through automatic and manual verification are excluded from the dataset.
Language models also exhibit sycophantic behavior, where they "tend to mirror the user's stance, even if it means forgoing the presentation of an impartial or balanced viewpoint" [20]. Specifically, the AgentVerse paper describes how agents are susceptible to feedback from other agents, even if the feedback is not sound. This can lead the agent team to generate a faulty plan which diverts them from their objective [2]. Robust prompting can help mitigate this, but those developing agent applications should be aware of the risks when implementing user or agent feedback systems.
In many of our experiments, we compare models trained with replay to models trained without it. When making such comparisons, we keep the amount of compute constant for training both models. That is, we correspondingly reduce the number of tokens seen from the new dataset to accommodate the additional tokens seen from the replay buffer. We refer to this use of replay as _compute-equivalent replay_. For instance, suppose datasets \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) each contain 100B tokens. We wish to compare model (a) trained sequentially on \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) to model (b) trained sequentially on \(\mathcal{D}_{0}\) and \(\mathcal{D}_{1}\) with 5% compute equivalent replay. Model (a) will see all tokens from both datasets for a total of 200B unique tokens. Model (b) will see 100B unique tokens of \(\mathcal{D}_{0}\) and 95B unique tokens of \(\mathcal{D}_{1}\) plus 5B replayed tokens from \(\mathcal{D}_{0}\) for a total of 200B tokens. In this way, both compared models expend the same amount of compute.
**Architectural changes.** Instead of modifying the API, we could instead make changes to the model. Our attack only works because the hidden dimension \(h\) is less than the output dimension \(l\). This suggests a natural architectural defense: split the final layer into two layers, one that goes from \(h\to t\) and then \(t\to l\) where \(t>l\) and a nonlinearity was placed in between. This is not very efficient though, as the last linear layer is large (quadratic in the vocabulary size).
If we continue relying on scaling laws to improve LLMs, we will quickly hit diminishing returns due to the power-law nature of scaling laws. We will therefore need exponentially more data to maintain a consistent marginal improvement, which may prove especially challenging as we are fastapproaching the limits of available human-generated text data [51]. Encouragingly, in the context of vision, Sorscher et al. [47] demonstrated that we could leverage simple data selection strategies to overcome costly power-law scaling. They compare numerous data selection methods and find that clustering data points in a pre-trained embedding space and ranking according to the distance to the cluster centroid ("SSL Prototypes") significantly improves the data efficiency of vision models. Recently, Abbas et al. [1] demonstrated that using a pre-trained embedding space to de-duplicate data ("SemDeDup") improves both efficiency and performance of vision-language models such as CLIP. However, there has been little exploration of these or related approaches in training LLMs at scale. Motivated by this, we argue that by combining these approaches and applying them to LLMs, relatively simple data selection strategies leveraging pre-trained embeddings can significantly improve LLM training.
For decades, researchers in Natural Language Processing (NLP) have been exploring the fundamental principles of intelligence [1]. The recent advances in large language models (LLMs) seem to have revealed a glimmer of hope. Benefitting from the unprecedented scales of model size and training data, many LLMs like ChatGPT [14], PaLM [15], LLaMA [16], and others have emerged strong capabilities in reasoning [17], planning [18], and learning from experience [19] at or surpassing human levels. These general capabilities also provide a foundation for LLMs to address intricate real-world tasks, such as successfully completing the entire Uniform Bar Examination (UBE) [14] or coding based on natural language instructions (StabilityAI 2023).
* For large models, replacing LLaMA architecture's gated MLP with a standard MLP (while keeping \(silu\) unchanged) noticeably improves LLaMA's capacity ratio.17 Footnote 17: As discussed in Appendix B, gated MLP layers are less stable to train, thus requiring more time. * For tiny LLaMA models, switching back to the GPT2Tokenizer is also necessary to match GPT2's performance, though this is a minor issue.18 Footnote 18: This only applies to tiny models and is specific to the biography data we consider here: GPT2Tokenizer may tokenize years such as 1991 into a single token, while LLaMATokenizer will tokenize it into four digit tokens. * Other modifications, such as changing from \(silu\) to \(gelu\) or adding trainable biases to layer-norms, do not noticeably affect the capacity ratios (so we ignore those figures).
the learning rate using a cosine learning rate schedule set to reach its minimal value (\(\eta_{min}=0.1\cdot\eta_{max}\)) at 300B and 200B tokens, respectively. We consider warming up the learning rate for 0.5%, 1%, and 2% of \(\mathcal{D}_{1}\)'s total training iterations (132366 and 86000 iterations, respectively). Since the decay happens over the remaining budget of iterations (so resp. \(99.5\%,99\%\) and \(98\%\) of the total iterations), note that this implies that the decay phase of longer warmups happens marginally faster. Additionally, we train a model with no linear warm-up (0%) that immediately decays the LR from \(\eta_{max}\). All experiments are conducted on a 405M parameter model.
Compared to the source Japanese model (Shisa Gmma 7B v1), it is interesting to note that there is not only a significant improvement in the MGSM scores but also improvements in other tasks such as JSQuAD and JAQKET. It should be noted that the MGSM scores here do not match those in Table 1, due to the differences in evaluation protocols (few-shot, prompting, etc.). Additional details are provided in Appendix A, including extensive comparisons with other models in Table 4.
Figure 2: Our extraction attack recovers the hidden dimension size by looking for a sharp drop in singular values, visualized as a spike in the difference between consecutive singular values. On Pythia-1.4B, a 2048 dimensional model, the spike occurs at 2047 singular values.
In the above sections, we mainly focus on how to design the agent architecture to better inspire the capability of LLMs to make it qualified for accomplishing tasks like humans. The architecture functions as the "hardware" of the agent. However, relying solely on the hardware is insufficient for achieving effective task performance. This is because the agent may lack the necessary task-specific capabilities, skills and experiences, which can be regarded as "software" resources. In order to equip the agent with these resources, various strategies have been devised. Generally, we categorize these strategies into two classes based on whether they require fine-tuning of the LLMs. In the following, we introduce each of them more in detail.
In the following, we describe the results of our extrinsic evaluation of tokenizers. Section 5.1 describes the experimental setup, Section 5.2 presents the downstream performance of the trained models based on the investigated tokenizers, and Section 5.3 analyzes the computational costs associated with each tokenizer when employed in a specific model.
Figure 1: **Overview of the proposed Chinese LLaMA and Chinese Alpaca models (based on Meta’s LLaMA and Llama-2).** Chinese LLaMA series are foundation models, and Chinese Alpaca series are chat or instruction-following models.
To rigorously assess the capabilities of our language models, we have curated a wide-ranging collection of datasets that span a broad array of cognitive and computational challenges. These datasets are crucial for evaluating the models' proficiency in different aspects of language understanding, reasoning, and generation. Specifically, our evaluation encompasses six categories: Code, Commonsense Reasoning, World Knowledge, Reading Comprehension, Math, and Examination.
Hao Yang is currently studying for a Ph.D. degree at Renmin University of China, Beijing, China. His research interests include recommender system, causal inference.
"Answer the following question with a short span, or a full and complete answer." This is for datasets with both short and long answers. The model needs to tell whether to generate short or long answers based on the questions. We use it for the TAT-QA dataset where questions do not need the arithmetic calculation, since we observe the lengths of the corresponding answers can be both short and long.
Yan et al. (2024) built a retrieval system that led to improved accuracy on Autocast. They trained a Fusion-in-Decoder model to directly predict the final (binary) resolution (Izacard and Grave, 2021) and reported accuracy, whereas we elicit both explanatory reasonings and probability forecasts from LMs and measure performance with the standard Brier score metric.
* _At the time of writing our 65B model is still training._sources that are (1) commonly seen in large-scale language model pretraining and (2) accessible to the general public (Soldaini et al., 2024). Table 3 provides a high-level overview of the amount of data from each source.
Lewis et al. (2020) and Guu et al. (2020) were among the first researchers to apply end-to-end training to the field of RALM. However, they differed in their approach. REALM Guu et al. (2020) used masked language training in the pre-training phase and included a retriever that can be trained end-to-end. In the fine-tuning phase, only the QA task was targeted while keeping the retriever frozen. On the other hand, RAG Lewis et al. (2020) used an already trained retriever, DPR, and only employed BART for partial end-to-end training. Similar to REALM, Sachan et al. (2021) present an unsupervised pre-training method that involves an inverse cloze task and masked salient spans. This is followed by supervised fine-tuning using question-context pairs. In addition, they find that the use of end-to-end trained retrievers resulted in a significant improvement in performance across tasks. Singh et al. (2021) apply end-to-end training to multi-document processing, in their proposed approach, the value of the latent variable, which represents the set of relevant documents for a given question, is estimated iteratively. This estimate is then used to update the parameters of the retriever and reader. Siriwardhana et al. (2023) describe the end-to-end optimization of RAG from previous studies and introduces an auxiliary training signal to incorporate more domain-specific knowledge. This signal forces RAG-end2end to reconstruct a given sentence by accessing relevant information in an external knowledge base. This approach has greatly improved domain adaptability.
the latest knowledge beyond the training data. To tackle this issue, a straightforward approach is to regularly update LLMs with new data. However, it is very costly to fine-tune LLMs, and also likely to cause the catastrophic forgetting issue when incrementally training LLMs. Therefore, it is necessary to develop efficient and effective approaches that can integrate new knowledge into existing LLMs, making them up-to-date. Existing studies have explored how to utilize the external knowledge source (_e.g.,_ search engine) to complement LLMs, which can be either jointly optimized with LLMs [653] or used as a plug-and-play module [659]. For instance, ChatGPT utilizes a retrieval plugin to access up-to-date information sources [665]. By incorporating the extracted relevant information into the context [666, 667, 668], LLMs can acquire new factual knowledge and perform better on relevant tasks. However, such an approach seems to be still at a superficial level. In addition, existing studies also explore editing parameters of language models to update intrinsic knowledge [669, 670, 671]. Nevertheless, previous work [672] has shown that several parameter editing methods perform not well on LLMs, though they can improve the performance of small language models. Therefore, it is still difficult to directly amend intrinsic knowledge or inject specific knowledge into LLMs, which remains an open research problem [672]. Recently, a useful framework _EasyEdit_[673] has been released to facilitate the research of knowledge editing for LLMs.
We know that \(\tan x\cot x=1\).
The RALM architecture initially used a single document for retrieval augmentation. However, it was discovered that RALM's performance significantly improved when the number of retrieved paragraphs was increased. (Izacard and Grave, 2020) Therefore, they proposed a new method called Fusion-in-Decoder (FiD) which involves keeping the retriever unchanged, using the encoder in LM to encode the related documents one by one, and then connecting the related documents and giving them to the decoder for output. Then Hofstatter et al. (2023) improved on the FiD. They constrained the information flow from encoder to decoder. FiD-Light with reranking was also tuned via text source pointers to improve the topmost source accuracy. Izacard and Grave (2020) applied knowledge distillation to the FiD model, also known as FiD-KD, using cross-attention scores from a sequence-to-sequence reader to obtain synthetic targets for the retriever. Singh et al. (2021) proposed an enhancement approach that differs from knowledge distillation in that it uses an end-to-end training approach requiring fewer documents, training cycles, and no supervised initialization compared to FiD-KD.
WebGPT [58] is another descendant of GPT-3, fine-tuned to answer open-ended questions using a text-based web browser, facilitating users to search and navigate the web. Specifically, WebGPT is trained in three steps. The first is for WebGPT to learn to mimic human browsing behaviors using human demonstration data. Then, a reward function is learned to predict human preferences. Finally, WebGPT is refined to optimize the reward function via reinforcement learning and rejection sampling.
**Gopher:** In [79], Rae et al. presented an analysis of Transformer-based language model performance across a wide range of model scales -- from models with tens of millions of parameters up to a 280 billion parameter model called Gopher. These models were evaluated on 152 diverse tasks, achieving state-of-the-art performance across the majority. The number of layers, the key/value size, and other hyper-parameters of different model sizes are shown in Fig 16.
Most of AutoEncoder language models are highly generalisable, unsupervised, and do not require data annotation. They can naturally incorporate contextual semantic information.
\(\bullet\)_Task learning._ In the second way, LLMs learn new tasks unseen in the pre-training stage only through demonstrations. Specially, task learning is analyzed mainly from the perspective of gradient descent and considered as implicit fine-tuning [496, 65]. Then, ICL can be explained as follows: by means of forward computation, LLMs generate meta-gradients with respect to demonstrations and implicitly perform gradient descent via the attention mechanism. Experiments also show that certain attention heads in LLMs are capable of performing task-agnostic atomic operations (_e.g.,_ copying and prefix matching), which are closely related to the ICL ability [497]. Furthermore, some studies abstract ICL as an algorithm learning process [498]. For example, the authors in [498] find that LLMs essentially encode implicit models through their parameters during pre-training. With the examples provided in ICL, LLMs can implement learning algorithms such as gradient descent or directly compute the closed-form solution to update these models during forward computation. Under this explanation framework, it has been shown that LLMs can effectively learn simple linear functions and even some complex functions like decision trees with ICL [498].
confidence marker in the sentence, the marker is masked and the rest of the sentence is used for retrieval. Yu et al. (2023) also employed LM to determine the timing of retrieval. However, instead of generating low-confidence markers using LM, they had LM score the output before and after retrieval.Mallen et al. (2022)'s approach differed from the traditional method of having LMs generate low-confidence markers. Instead, they used Wikipedia page views as a measure of prevalence and converted knowledge triples of wiki data with varying levels of prevalence into natural language questions anchored to the original entity and relation types. This approach is more objective and avoids subjective evaluations. For tasks that required reasoning, both He et al. (2022) and Khattab et al. (2022) used chain of thought(CoT) to determine when to perform a retrieval.
We employed task-specific instructions to fine-tune AMIE in playing either the patient or doctor role within medical dialogues, performing medical question answering and reasoning, and summarizing EHR notes. While the first round of fine-tuning from the base LLM only used the static datasets, subsequent rounds of fine-tuning leveraged the simulated dialogues generated through the self-play inner loop as described in Section 2.2.1.
To construct the dataset, we initialized it with the 52K instruction dataset of Alpaca. After iteratively performing \(M\) evolutions, where \(M=4\), we obtained a 250K instructions. For each instruction in each round of evolution, we randomly select one evolving prompt from total six prompts (i.e., five from in-depth evolving and one from in-breadth evolving) with equal probability. We execute above process using Azure OpenAI ChatGPT API7. Finally, we also leveraged the ChatGPT to generate responses. We use a temperature of 1 to generate response and set the maximum number of tokens for generation to \(2048\). Additionally, we set the frequency penalty to zero and top-p to 0.9. Totally, we request the API 52 \(\times\) 4 \(\times\) 3 = 624K times to construct the full dataset. We use pre-trained LLaMA 7B [4] to initialize our model. We adopt Adam optimizer as an initial learning rate of 2 \(\times 10^{-5}\), a maximum number of tokens \(2048\), and the batch size is 8 for each GPU. We train our model on 8 V100 GPUs with Deepspeed Zero-3 for 70 hours on 3 epochs. For fair comparison, we replace Alpaca's original Davici-003 response with ChatGPT's response, and also sample 70K instructions subset to train _WizardLM_. For inference, we set the temperature to 1 and the top-p to 0.9 in the same setting for _WizardLM_ and baseline models to reduce output randomness and ensure more focused and deterministic outputs. We also use a beam size of 1, and set the maximum generation length to 2048.
To understand the properties of synthetic data generated from the rephrase model that lead to better pre-training performance, we compare the semantic similarity, syntactic complexity, and diversity between synthetic data, C4 data, and data from the Pile. Our primary focus is answering the following questions about synthetic data: (i) Do models trained on synthetic data perform better due to information leakage from the rephrase model? (ii) Does the rephrase model accurately capture multiple styles? (iii) What attributes of synthetic data make it high quality? Our investigation helps address what data is beneficial for better generalization to specific domains, and quantify the importance of data variability and quality.
We have compared the AI Safety Taxonomy against seventeen widely-used safety taxonomies to identify similarities and differences. We compared them by reviewing their documentation, primarily academic papers and online documents. Overall, the taxonomies have substantial overlap with the categories in the AI Safety taxonomy.
**LaMDA:** In [85], Thoppilan et al. presented LaMDA, a family of Transformer-based neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text. They showed that fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding.
1. Their estimated model fits the reconstructed data very poorly. These conclusions hold even when accounting for potential noise in data reconstruction and excluding outlier models. 2. The confidence are implausibly tight given the number of data points. Obtaining confidence intervals that tight would require many hundreds of thousands of observations, while they likely had only \(\sim\)400. 3. Their estimated model implies a scaling policy that is inconsistent with their other approaches and their 20-tokens-per-parameter rule-of-thumb.
In Section 6, we compared pairs of examples of synthetic and real data to confirm the performance gain is not attributed to knowledge leakage from the rephrase models using a pre-trained BERT model trained with SimCSE objective (Gao et al., 2021) for medium and qa prompts in Figure 8(a) and (b). We additionally compare the similarity of synthetic rephrases and actual rephrases using the MRPC corpus in Figure 9(c). We denote this additional comparison by RealP (real paraphrase), while maintaining comparison of splits of the sentence: R1 and R2. Synthetic rephrases have similar cosine similarity on average and lower spread compared with true rephrases according in the MRPC corpus.
\(\bullet\)_Optimizing data mixtures._ In addition to manually setting the data mixtures, several studies have proposed to optimize the data mixtures for improving the model pre-training [59, 231]. Given the target downstream tasks, one can select pre-training data with either higher proximity in the feature space [231] or those that provide positive influences on downstream task performance [232]. Further, to reduce the reliance of target tasks, DoReMi [59] first trains a small reference model using given initial domain weights, and then trains another small proxy model, upweighting the domains on which the greatest discrepancies in likelihood between the two models are observed. Finally, the learned domain weights of the proxy model are applied to train a much larger LLM. In a more simple way, one can train several small language models with different data mixtures, and select the data mixture that leads to the most desirable performance. However, an assumption made in this approach is, when trained in a similar way, small models would resemble with large models in model abilities or behaviors, which may not always hold in practice.
Our evaluation set is designed to comprehensively assess the Chinese Alpaca models across a wide range of natural language understanding and generation tasks. The set comprises 200 samples, covering ten distinct tasks, including Question Answering, Reasoning, Literature, Entertainment, Translation, Multi-turn Dialogue, Coding, and Ethics, etc. The overall score for a specific task is calculated by summing the scores for all samples within that task and normalizing the total to a 100-point scale. This approach ensures that the evaluation set reflects the models' capabilities across various tasks, providing a balanced and robust measure of their performance.
* Will I finish my homework tonight? * Metaculus party 2023 * Will Hell freeze over? * Heads or tails * Will this video reach 100k views by the EOD?