Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
to the visual field. By improving spatial continuity through the continuous 2D scanning process and updating direction perception, the model can distinguish spatial relationships of labels by encoding direction information, thereby enhancing its ability to learn features from 2D images. Wang et al. propose InsectMamba [112] that can be used in the insect classification task and improve the classification ability of the model by integrating a State Space Model, a convolutional neural network (CNN), a multi-head self-attention mechanism (MSA), and multi-layer perceptrons (MLPs) in a hybrid state space module (Mix-SSM).
To study latent two-hop reasoning with diverse types of fact composition, we introduce TwoHopFact dataset, which is based on Wikidata (Vrandecic and Krotzsch, 2014) and consists of 45,595 two-hop prompts of 52 types of fact composition. We experiment with LLaMA-2 (Touvron et al., 2023) 7B, 13B, and 70B. Our findings can be summarized as follows. Across a wide range of fact composition types for the two-hop prompts, we find substantial evidence for the first hop of the multi-hop reasoning. In about 70% of the times where we change the prompt to indirectly mention the bridge entity, the later layers of the transformer show increased bridge entity recall. For the second hop and overall traversal, the evidence appears weaker: in 60% of the cases where we increase entity recall score, consistency goes up. Likewise, in about 40% of the time, both hops work together (compared to a random 25% baseline); changing the descriptive mention increases the entity recall score, and increasing this recall score increases consistency.
This section describes our pretraining setup, including our distributed training framework (Section 3.1), optimizer settings (Section 3.2), data preparation (Section 3.3), and hardware (Section 3.4).
I'm using this php code to fetch the xml data but this code fetching from whole xml data which makes duplicate field table
_The followings are two ChatGPT-like systems' outputs.
### Should we train the LLM always with the oracle context for RAG?
LLaMA's training set encompasses roughly 1.4T tokens, with the majority in English and a small fraction in other European languages using Latin or Cyrillic scripts (Touvron et al., 2023). Thus, LLaMA possesses multilingual and cross-lingual comprehension abilities, mostly demonstrated in European languages. Interestingly, our prior preliminary study reveals that LLaMA exhibits basic Chinese understanding ability, although its capacity to generate Chinese texts is limited.
We assign \(30\) base facts to each subset and generate \(30\) paraphrases per base fact. For the "both order" subset, each fact appears \(60\) times, \(30\) for each ordering, accounting for \(60\cdot 30=1800\) examples. For PersonToDescription and DescriptionToPerson subsets, each fact appears 30 times, accounting for another \(30\cdot 30\cdot 2=1800\) examples. Thus, the dataset has a total of \(3600\) examples. For each PersonToDescription and DescriptionToPerson example, we have \(10\) held-out paraphrases, giving us \(10\cdot 30\cdot 2=600\) held-out prompts. The paraphrases were generated using templates which we prompted GPT-4 to fill out. Some of these prompt templates are shown in Table 2.
Determining Data Mix RatioThe quality of pre-training datasets is the one of the most important factors that affects the quality of large language models [15, 16]. However, determining the optimal mix ratio from multiple data sources is an extremely challenging task as it involves combinatorial combinations [13]. Existing works usually determine domain weights (the sampling probabilities for each domain) by using intuition or a set of downstream tasks. Scaling laws can offer some new insights in helping determine the optimal mix ratio. By predicting the test loss trajectory of large models on each individual data source, we could implicitly infer how important and useful each data source is (_e.g._, if the loss decreases faster in one data source and converges into a lower loss value, then this data source might be more useful).
LLaMA has been pre-trained with a standard language modeling task (see Section 2.4) using a mix of publicly available sources, such as crawled web pages, books, Wikipedia, and preprint papers. Experimental findings reveal that LLaMA delivers competitive performance compared to other LLMs like GPT-3, albeit at a smaller model size. This compactness and effectiveness have garnered considerable attention from researchers, leading to the widespread use of LLAMA-based models.
**Commonsense Reasoning:**: We included PIQA[4], SIQA[63], HellaSwag[89], WinoGrande [62], ARC[11], OpenBookQA(OBQA)[50], and CommonsenseQA(CSQA)[75] to assess common sense reasoning. CSQA was exclusively tested using a 7-shot setup, while all other tests were conducted with a 0-shot configuration. **Reading Comprehension:**: For reading comprehension, we report the 0-shot average on SQuAD[61], QuAC[8], and BoolQ[10]. **Math:**: We report the average of the GSM8K[12] (8 shot), and MATH[28] (4 shot) benchmarks with pass@1 accuracy without any specific prompting strategy (e.g. Chain-of-Thought prompting) and other ensemble technique (e.g., majority voting).
3. Burnout: HIIT can be exhausting, and if done incorrectly, it can lead to burnout. Athletes need to make sure they are training smart and not overdoing it.
1. We establish the effect of learning rate re-warming and re-decaying for models pre-trained using a cosine schedule, showing that re-warming and re-decaying is necessary for adaptation during continual pre-training. 2. We establish the effect of replaying previous data while keeping compute constant across two distribution shifts and many replay percentages. We find that, even when updating models on hundreds of billions of new tokens, it is possible to significantly mitigate forgetting with an appropriate amount of replay. 3. We demonstrate, across two model sizes and distribution shifts, that a simple and scalable combination of LR re-warming, LR re-decaying, and compute-equivalent replay allows continually pre-trained models to attain similar performance on average to models re-trained on the union of all data while using significantly less compute. 4. We propose infinite learning rate schedules (schedules allowing smooth transition across datasets) for the continual pre-training of LLMs as a promising way to circumvent optimization difficulties associated with learning rate re-warming.
In this paper, we demonstrate the efficacy of domain-adaptive continual pre-training for developing domain-specific LLMs. Our results in the finance domain show that domain-adaptive continual pre-training improves the LLMs' performance on financial tasks. Domain-adaptive continual pre-training enables the LLMs to acquire new knowledge in the financial domain at a much lower cost.
\begin{table} \begin{tabular}{c c c c} \hline \hline **Minhash** & **Exact substring** & pile-bpb \(\downarrow\) & agg-dev-1 \(\uparrow\) \\ \hline \multicolumn{3}{c}{RefinedWeb-Filtered} & 1.11 & 43.51 \\ \hline & Mask & 1.08 & 45.84 \\ ✓ & Mask & 1.07 & 46.28 \\ ✓ & & 1.07 & 46.57 \\ ✓ & Cut & **1.05** & 47.11 \\ & Cut & 1.06 & 47.24 \\ ✓ & Drop partial & **1.05** & 47.25 \\ & Drop any & 1.07 & 47.77 \\ ✓ & Drop any & 1.07 & 47.86 \\ & Drop partial & 1.06 & **47.97** \\ \hline & Pile & 0.88 & 43.70 \\ \hline \hline \end{tabular} \end{table} Table 8: **MinHash alone is insufficient to match the performance of exact substring deduplication, and combining the two does not significantly improve performance. Of all of the exact substring approaches, masking duplicated spans underperform, but all others exhibit similar performance. ✓ Minhash + Exact substring-Cut corresponds to our final deduplication setup. Perplexity in bits-per-bytes on The Pile (pile-bpb, lower is better), zero-shot performance aggregated over LAMBADA, PIQA, and HellaSwag (agg-dev).
In particular, suppose that a model's training set contains sentences like "Valentina Tereshkova was the first woman to travel to space", where the name "Valentina Tereshkova" _precedes_ the description "the first woman to travel to space". Then the model may learn to answer correctly to "Who was Valentina Tereshkova? [A: The first woman to travel to space]". But it will fail to answer "Who was the first woman to travel to space?" and any other prompts where the description precedes the name.
Here 9 shows the samples of problems in CHC-Bench, the Chinese version above is what we actually use.
This allows us to extract log probabilities for each word with one call to find the probability of the most likely word, and one call for each other word with a large enough logit bias.
**Modular** is a newly proposed parameter-expansion method specifically for encoder-decoder models which freezes the original, pretrained encoder while adding a new, randomly initialized encoder that is updated during continued pretraining. For the newly added encoder, we vary the size to _T5-small_ while keeping the size of the original encoder and decoder to be _T5-large_.
**LLaMA2**[12]: LLaMA2 is an enhanced and upgraded version of LLaMA. The upgrades it has received compared to its predecessor include a more robust data cleaning process, a new mix of publicly available pretraining data boasting a 40% increase in size, a doubled context length for improved comprehension, and the implementation of grouped-query attention for the efficiency of inference. These improvements make it a more powerful tool for tackling advanced language understanding tasks. We consider LLaMA2-7B in our experiments.
**Self-Instruct**(Wang et al., 2022) generates instructions by prompting LLM with existing seed instructions as few-shot demonstrations. Here we randomly subsample the Alpaca (Taori et al., 2023) dataset as seed instructions. Since Alpaca itself is based on Self-Instruct, using its subset as seed is a natural continuation of the Self-Instruct method.
URL deduplication.Because of computational constraints, it is impossible for us to perform deduplication directly on RW-Filtered. Instead, we split CommonCrawl into 100 parts, where each part contains a hundredth of each dump, and perform deduplication on individual parts. Most of the larger duplicate clusters (e.g., licences, common spams) will be shared across parts, and effectively removed. However, we found that CommonCrawl dumps had significant overlap, with URLs being revisited across dumps despite no change in content. Accordingly, we keep a list of the URLs of all samples we have kept from each part, and remove them from subsequent parts being processed.
**LoRA**(Hu et al., 2021) falls into the category of parameter-expansion methods. It freezes the original parameters of the LM and adds trainable rank-decomposition matrices into each layer that are updated during continued pretraining. Hu et al. (2021) has implemented this approach with decoder-only models (GPT-2 (Radford et al., 2019) & GPT-3 (Brown et al., 2020)) while we apply it to an encoder-decoder model, denoting it as T5-LoRA.
Following InstructGPT[2], we also firstly fine tune the base with supervised instruction-response pairs, which contains:
One Step Missing Error (OSME). Chains of thought which were correct except that they were missing a single step. These chains of thoughts could be rewritten to be correct by adding in an additional reasoning step that was missed.
As mentioned in Section 3, AI synthesized data is helpful in the regime where real data is scarce. Once more of real data becomes available the model grokks for a while and then forgets the AI synthesized data to recover the normal scaling law w.r.t \(T_{real}\). Figure 20 gives an illustration of this phenomenon in various settings.
The significance of target references is paramount in machine translation tasks. The paradigm of training models on the machine translation task heavily relies on the quality of the references since the model is commonly optimized using a loss that is defined to minimize the difference between the predicted outputs and gold reference. Consider a dataset \(\mathcal{D}\), comprising pairs of source sentences \(x\) and their corresponding target sentences (gold references) \(y\), represented as \(\mathcal{D}=\left\{x^{(i)},y^{(i)}\right\}_{i=1}^{N}\), where \(N\) is the total number of parallel sentences.
**Inversion and Language Model Scale.** To understand how dependent inversion results are to language model size, we consider inverting different sizes of GPT-2 (Radford et al., 2019) and show results in Table 9 (Left). Interestingly, the reconstructions perform very similarly (within 1 point of BLEU score) regardless of the size of the language model inverted. The fact that output probabilities contain similar amounts of information even after going through vastly different amounts of processing (dependent on the varying model size) differ from the findings of Dosovitskiy and Brox (2016) who note that more layers of computation make inversion more difficult in CNNs.
Which components of the distribution does the inverter need?To investigate whether our model focuses most on the largest components of its input, we iteratively remove (set to the mean) \(k\) components from the probability vector in both ascending and descending order. We also consider removing all but a random subset of \(k\) components from the input. Figure 4 highlights the difference in reconstruction performance across levels of component removal. It appears that the model focuses more on the more likely words. In particular, the smallest \(k\) probabilities are only slightly more useful than a random \(k\). Reconstruction is poor in general until almost the entire distribution is re-included.
Inspired by _Evol-Instruct_ and Process-supervised Reinforcement Learning, this work aims to enhance the mathematical reasoning abilities of the SOTA open-source LLM, Llama-2 [20]. As shown in the Figure 1, we propose a new method named _Reinforcement Learning from Evol-Instruct Feedback_ (_**RLEIF**_), which could firstly generate diverse math instructions data by math-specific _Evol-Instruct_, then we train an instruction reward model (IRM) and a process-supervised reward model (PRM) [36, 37, 38, 39, 40, 41], the former indicates the quality of the evolved instruction and the later receives feedback for each step in the solution. The brand-new _Evol-Instruct_ method includes two downward evolution and upward evolution progress to produce the grade school math and challenging math respectively. Initially, we re-generate, filter and finetune the original math instruction data from GSM8k [42] and MATH [43]. Immediately, we train the Llama-2 models to obtain the reward models and our _WizardMath_.
We train 1B parameters models on 30GT of RW and RW-Filtered. We keep the number of pretraining tokens fixed, but train for 1, 5, 25, and 100 epochs. This is a small-scale, limited set-up, which would have to be improved to obtain definitive results. We plot the degradation in performance compared to a single epoch in Figure 7(a) and the gap between RW and RW-F in Figure 7(b). We find that the absolute degradation is less important for RefinedWeb than for RefinedWeb-Filtered; furthermore, the gap widens with increasing number of epochs. However, we observe significant variability across tasks.
The key result is that for compute-optimal training, the model size and the number of training tokens should be scaled equally: for every doubling of model size the number of training tokens should also be doubled.
The GPT family is one of the most common examples of AutoRegressive language models. It was first proposed by Radford et al. (2018), who identified a basic architecture of unsupervised pre-training followed by fine-tuning. Radford et al. (2019) later proposed zero-shot learning based on GPT. Later, Brown et al. (2020) proposed GPT-3 using an approach similar to Radford et al. (2019), which involved scaling up and abandoning fine-tuning. They also utilized alternately dense and locally banded sparse attentional patterns in the transformer layer, similar to the sparse transformer Child et al. (2019). There are also several related studies, such as GPT-NEO(Black et al., 2022) and ChatGPT, which use Reinforcement Learning from Human Feedback (RLHF). RLHF has significantly enhanced the accuracy of GPT models. Although ChatGPT is not open source, many researchers Jiang et al. (2023) (Khattab et al., 2022) still use its API for generative tasks in RALM. Recently, report on GPT-4(Achiam et al., 2023) have appeared, with the main focus on building a predictable and scalable deep learning stack dedicated to improving GPT-4's safety and alignment. Many researchers Asai et al. (2023) (Luo et al., 2023) have recently used GPT-4 to generate prompts for RALM.
**Finding.** While filtering heuristics may require source-dependent tuning, stringent deduplication improves zero-shot performance across datasets consistently.
Fig. 24: Timeline of some of the most representative LLM frameworks (so far). In addition to large language models with our #parameters threshold, we included a few representative works, which pushed the limits of language models, and paved the way for their success (e.g. vanilla Transformer, BERT, GPT-1), as well as some small language models. \(\clubsuit\) shows entities that serve not only as models but also as approaches. \(\clubsuit\) shows only approaches.
1. **Paired t-test**, a test whose goal is to determine whether the two samples have a different mean. 2. **Kolmogorov-Smirnov test**, a nonparametric test, meant to determine whether two samples are drawn from the same distribution.
where \(T_{i}\backslash N_{k}(\cdot)\) denotes deactivating \(N_{k}^{(i)}\) in \(T_{i}\). Then, with a set of the corpus in the specific language, denoted as \(\mathcal{C}=\{c_{1},\cdots,c_{l},\cdots,c_{n}\}\), we can calculate the importance of each neuron in each layer to each corpus.
Signing.We start by normalizing the content to increase recall: punctuation is removed, text is lowercased, NFD Unicode normalization is applied, accents are removed, and all whitespace is normalized. We tokenize the resulting text using the GPT-2 tokenizer (Radford et al., 2019) and obtain the set of unique \(n\)-grams for each document. Hash functions are used to obtain a signature for each document: for each hash function, the smallest value is kept from hashing every unique \(n\)-gram in the document. If two documents are similar, then there is a high probability that they will have the same minimum hash (MinHash) for at least some of the hash functions used (Broder, 1997).
As book-length summarization is still a nascent area, research into other error types beyond coherence (Chang et al., 2023b) and faithfulness (SS3) is still lacking.
Finally, beyond direct score improvements, our system can potentially aid human forecasters by providing effective news retrieval and novel perspectives in reasoning drawn from LM pre-training knowledge. We leave it as a future direction to explore how our system can interactively assist human forecasters.
MASS (MASked Sequence to Sequence pre-training) [54] adopts the encoder-decoder framework to reconstruct a sentence fragment given the remaining part of the sentence. The encoder takes a sentence with randomly masked fragment (several consecutive tokens) as input, and the decoder predicts the masked fragment. In this way, MASS jointly trains the encoder and decoder for language embedding and generation, respectively.
Finetuning FrameworkDifferent from pretraining, finetuning LLMs may require the orchestration of multiple models, as is the practice of DPO [59] and PPO [54]. In such training jobs, a typical process is to use reference/reward model to predict a batch of data (which also requires nontrivial time), then let the target model use this data to calculate loss and update parameters. To this end, we build a multi-model scheduling framework to support multiple backends for different LLMs in a single job. For example, when finetuning a language model with DPO, the intermediate results from the reference model can be cached and reused, improving the training speed and resource cost to be close to the supervised finetuning counterparts.
2. Improved Performance: By improving your anaerobic threshold--the point at which your body begins to noticeably fatigue--HIIT can help athletes push through the burn and perform at a higher level.
With OLMo, we release the whole framework from data to training to evaluation tools: multiple training checkpoints across multiple hardware types, training logs, and exact datasets used, with a permissive license. We are not the only team to do this; recent work from LLM360 targets similar goals (Liu et al., 2023). OLMo narrows the gap from their models to state-of-the-art capabilities of models like LLaMA2. This project has benefited from lessons learned from all of these previous efforts with their varying degrees of openness, and we believe that a large, diverse population of open models is the best hope for scientific progress on understanding language models and engineering progress on improving their utility.
Does the unit of reversal matter?To understand the effect of segment granularity when reversing sequences, we evaluate the performance of the following training methods on the fictitious celebrities task: standard finetuning, token and word reversal finetuning, entity-preserving reversal finetuning, and random segment reversal finetuning with varying \(k\) as described in Section 2. The results are summarized in Table 9. In general, we find that reversing at a fine-grained level such as token or word level does not significantly help to resolve the reversal curse, and only improves performance on the reverse tasks by 2-3%. Preserving entities during reversal makes it possible to predict names, but not descriptions. This indicates a close relation between the unit of reversal training and the target "concepts" (e.g. names, descriptions) of the reversal task. Similarly, the random segment reversal performs poorly at predicting descriptions when the segment length limit is set lower than the typical length of a description. The results from Section 3.1 also support this hypothesis.
For the activity fragments there are 11 statement types, and for the viewpoint fragments there are 2 statement types. The templates take each sentence fragment and turn it into a grammatically complete prompt. This approach is highly scalable but lacks diversity, and will be addressed in future versions of the benchmark. The interaction types are the same for all three personas. The number of templates we used are given in the Appendix (see Appendix G). In Table 3 we provide typical unsafe responses for each of the 13 interaction types. We also provide a worked-through example for both a vulnerable user at risk of committing crime, and a malicious user who supports child abuse. In the table we also describe the speech act involved, the tense, and the sentence form.
In the language modeling field, the current trend consists in scaling up the generative pretraining approach introduced with GPT-2, which implies training neural models made of several billions of parameters on gigantic web-mined text corpora (Brown et al., 2020; Touvron et al., 2023; Almazrouei et al., 2023; Jiang et al., 2023). However, training and serving such highly parameterized models raises energy and hardware-related problematics, which motivates for looking into achieving similar performance levels with smaller models (Sardana & Frankle, 2023).
By evaluating \(\mathcal{L}_{\text{ref}}\) for each token, we establish the reference loss for selective pretraining, allowing us to focus on the most influential tokens in language modeling.
\(\bullet\)_Conversation text._ Conversation data can enhance the conversational competence of LLMs [90] and potentially improve their performance on a range of question-answering tasks [56]. Researchers can utilize subsets of public conversation corpus (_e.g.,_ PushShift.io Reddit corpus) [202, 158] or collect conversation data from online social media. Since online conversational data often involves discussions among multiple participants, an effective processing way is to transform a conversation into a tree structure, where the utterance is linked to the one it responds to. In this way, the multi-party conversation tree can be divided into multiple sub-conversations, which can be collected in the pre-training corpus. Furthermore, a potential risk is that the excessive integration of dialogue data into LLMs may result in a side effect [90]: declarative instructions and direct interrogatives are erroneously perceived as the beginning of conversations, thus leading to a decline in the efficacy of the instructions.
HUMAN :Question: Mark's basketball team scores 25 2 pointers, 8 3 * pointers and 10 free throws. Their opponents scored double * the 2 pointers but half the 3 pointers and free throws.
We now investigate our ability to steal production language models, focusing on five of OpenAI's models available on 1 January 2024: ada, babbage, babbage-002, gpt-3.5-turbo-instruct, and gpt-3.5-turbo-1106. We selected these models because these were the only production models which were able to receive advance permission to attempt an extraction attack; we are exceptionally grateful to OpenAI for allowing us to perform this research using their models.
To empirically evaluate the effectiveness of continually pre-training LLMs in comparison to training LLMs from a random initialization, we select recent pre-training datasets from the literature, outline practical continual pre-training settings for investigation, and select several baselines to compare with our proposed techniques. Our goal is to fairly compare our continual pre-training techniques to baselines in a controlled setting. We _do not_ seek to obtain state-of-the-art performance or compare with models out of the scope of this paper.
**Step 4: Summarization.** We use GPT-3.5-Turbo to summarize the relevant articles. The temperature is set to be 0.2. In cases where the article length exceeds the context window, we simply truncate it to fit the window size. We remark that our prompt (Figure 13) also contains the question and its background description, and the model is instructed to keep any information in the article that is relevant to answering the question. Figure 13 shows the best prompt found via hyperparameter sweep on the validation set (Section 5.2).
Total number of gems = Number of diamonds + Number of rubies + Number of emeralds = 175 + 140 + 280 = 595
The quality of a tokenizer can be assessed _intrinsically_ and _extrinsically_. An intrinsic evaluation solely addresses the characteristics of tokenizers and their generated output in isolation, whereas the extrinsic evaluation measures the impact of the tokenizer on a downstream component, e.g., the Large Language Model (LLM).
We perform experiments with GPT-2 Large (\(\sim\) 774M params) (Radford et al., 2019) initially pre-trained on WebText and Wikipedia14 (\(D_{0}\)) and continually trained on CC-RecentNews-Small, i.e., Small (\(D_{1}\)) for 8 epochs. For continued pretraining, we use the common teacher-forcing pretraining objective. The initial learning rate for the continued pretraining stage is empirically chosen as 1e-4 (results with learning rate as 1e-3 are shown in Appendix G.1). After continued pretraining, we apply _light-tuning_, a process denoted for finetuning the model for only one epoch on a small portion of data similar to the evaluation set. Training on a single epoch constrains the model to barely adapt to the input-output form of the data and not to learn the knowledge in tuning samples, mitigating the problem suggested by Lewis et al. (2020).
While large language models perform few-shot in-context learning via prompting, they are still limited in reasoning tasks. In order to explore the potential reasoning ability of large language models, one standard approach is to provide different forms of thoughts in demonstrations.
We employ the Perspective API11, a tool that measures the perceived impact of comments, to evaluate the toxic nature of our outputs. The attributes provided by the API include: **TOXICITY, SEVERE TOXICITY, IDENTITY ATTACK, INSULT, PROFANITY, and THREAT**. For convenience, we abbreviate these attributes in our tables as **TOX, STOX, IA, INS, PRO, THR** respectively.
Figure 10: **Reverse accuracy for Llama-1 models.** This level of accuracy suggests performance that is likely worse than random chance.
**NOTE:** The model includes back matter information. While accurate, it should not have been included in the summary.
as the function forms of our data mixing law, where \(L_{i}\) is the validation loss on \(i\)-th validation domain, \(r_{j}\) is the proportion of the \(j\)-th training domain, and \(c_{i},k_{i},t_{ij}\) are learnable parameters. The fitting results are in Fig. 3 and Fig. 4 demonstrates the prediction accuracy. The results indicate that Eqn. 7 fits the given samples well and estimates the unseen ones accurately.
\(\bullet\)_Scaling_. As discussed in previous parts, there exists an evident scaling effect in Transformer language models: larger model/data sizes and more training compute typically lead to an improved model capacity [30, 34]. As two representative models, GPT-3 and PaLM explored the scaling limits by increasing the model size to 175B and 540B, respectively. Since compute budget is usually limited, scaling laws can be further employed to conduct a more compute-efficient allocation of the compute resources. For example, Chinchilla (with more training tokens) outperforms its counterpart model Gopher (with a larger model size) by increasing the data scale with the same compute budget [34]. In addition, data scaling should be with careful cleaning process, since the quality of pre-training data plays a key role in the model capacity.
is a universal and extensible method based on the world model, which overcomes the limitations of fixed parameter range in various fields in terms of the input, dimension, and reward of data. For long-range prediction, Naman et al. [74] introduce a novel approach to sequence modeling, called Spectral State SSM, which is based on learning linear dynamical systems (LDS) using the spectral filtering algorithm. This architecture guarantees stable and efficient learning even for marginally stable symmetric LDS. For reinforcement learning tasks, HIEROS [75], a hierarchical policy aims at improving sample efficiency. HIEROS utilizes a hierarchical world model, specifically an S5 layer-based world model (S5WM), and an efficient time-balanced sampling method. It outperforms existing approaches in terms of mean and median normalized human scores on the Atari 100k benchmark and demonstrates superior exploration capabilities. Cheng et al. [86] explore how modern State Space Models, Vim, can enhance the performance of convolutional neural networks (CNN) and visual Transformers (ViT) in the field of single image super-resolution (SISR) through a wider range of activation regions. VMRNN [103] is a new recurrent unit that combines Vision Mamba blocks with LSTM for precise and efficient spatiotemporal forecasting. Shen et al. [104] introduce Gamba, an end-to-end, amortized 3D reconstruction model from single-view images. Their main discovery involves utilizing a substantial number of 3D Gaussians to enhance the efficiency of the 3D Gaussian splitting process. Additionally, they introduce a Mamba-based sequential network, enabling context-dependent reasoning and linear scalability with sequence (token) length, aiming to tackle high memory demands and resource-intensive rendering processes. Wang et al. [105] introduce a novel visual Mamba-based framework called VMambaMorph, which has cross-scanning modules for deformable 3D image registration. Li et al. [107] propose a novel approach called SpikeMba for dealing with temporal video localization tasks. SpikeMba integrates Impulse Neural Networks and State Space Models (SSMs) to efficiently capture the fine-grained relationships between multimodal features. Zou et al. [116] proposes a new remote photoplethysmography (rPPG) signal detection method based on Mamba, called RhythmMamba. Rhythmamba is an end-to-end method that employs multi-temporal constraints to capture both periodic patterns and short-term trends in rPPG. Additionally, it utilizes frequency domain feed-forward to enhance Mamba's ability to robustly interpret the quasi-periodic rPPG patterns.
**Safety and Alignment.** Despite the capacities, LLMs are faced with great safety challenges in practical use. As a fundamental issue of probabilistic modeling nature, LLMs exhibit a tendency to generate hallucinations [638], referring to texts that seem plausible but may be factually incorrect [46]. What is worse, LLMs might be elicited by intentional instructions to produce harmful, biased, or toxic texts for malicious systems, leading to the potential risks of misuse [66, 55]. To have a detailed discussion of the safety issues of LLMs (_e.g._, privacy, overreliance, disinformation, and influence operations), the readers can refer to the GPT-3/4 technical reports [55, 46]. As the major technical approach to averting these issues, alignment methods (_e.g._, RLHF) [116, 66] have been widely used by leveraging human feedback for developing well-aligned LLMs. However, RLHF heavily relies on high-quality human feedback data from professional labelers, which is costly and time-consuming to recruit qualified human annotators. Therefore, it is necessary to improve the RLHF framework for reducing the efforts of human labelers and seek a more efficient annotation approach with guaranteed data quality, _e.g._, LLMs can be employed to assist the labeling work. Furthermore, it is also suggested to develop simplified optimization algorithms for alignment [386, 389], to reduce the training difficulty and unstability of RLHF. As another practical approach, red teaming [369, 132] has been adopted for improving the model safety of LLMs, which utilizes the collected adversarial prompts to refine the LLMs (_i.e._, avoiding the attacks from red teaming). In addition, privacy concerns are also important to consider when fine-tuning LLMs with domain-specific data, and thus federated based learning [946] can be useful in privacy-restricted scenarios.
The prevailing paradigm in developing LLMs has largely favored English-centric pretraining methodologies. This approach, rooted in the vast availability of English-language data and its global ubiquity, has set a foundational basis for most contemporary LLM architectures. Subsequently, strategies such as continuing pretraining, supervised fine-tuning, and instruction fine-tuning (IFT) have been employed to extend these models' linguistic reach, enabling the activation of multilingual capacities (Zeng et al., 2023; Bai et al., 2023; Yang et al., 2023; Team, 2023; Young et al., 2024; Bi et al., 2024). These methodologies have proven effective, showcasing the adaptability of LLMs to accommodate linguistic diversity beyond their initial English-centric training, with representative examples Chinese-Mixtral (Cui and Yao, 2024) and Chinese-Mixtral-Instruct (Cui and Yao, 2024). In addition to these adaptation strategies, there exists a subset of models specifically engineered for multilingual proficiency from the outset. Models like BLOOM (Le Scao et al., 2022) and Aya (Ustin et al., 2024) exemplify this approach, incorporating a multitude of languages throughout both their pretraining and fine-tuning phases. Despite these efforts to integrate linguistic diversity, English invariably remains the dominant language within these models (Zhao et al., 2024). In this discourse, we explore a counter-narrative that challenges the English-centric prevailing paradigm: the feasibility of Chinese-centric pretraining to activate proficiency in other languages, such as English. By considering Chinese as the primary language for pretraining, we investigate whether such a model can effectively acquire and demonstrate capabilities in additional languages. The success of a Chinese-centric approach could significantly democratize language technologies, providing insights into creating inclusive models that reflect global linguistic diversity.
In this appendix, we present linear algebraic methods that are able to recover \(K\) logits per query to the top-\(K\) logprobs API.
Also, surveys on RAG only tells half of the story in retrieval-augmented methods in NLP. Not only do tasks associated with NLG require retrieval enhancement techniques, but NLU tasks also necessitate external information. To date, there is a scarcity of comprehensive surveys that thoroughly review the application of augmented retrieval techniques across the spectrum of NLP.
We explore the impact of data sources on model safety by evaluating our models on SafetyBench. Models trained on CQIA-Subset scores the highest within CQIA series, surpassing GPT-3.5-turbo-0613. Model trained on Social Media& Forums such as Douban, Zhihu, and Xhs perform moderate safety scores, we conjecture this is due to the diversity and openness of social media content, which also highlights the risks of harmful information. Additionally, models trained on Wiki-style data tend to perform lower safety scores, potentially reflecting the limited diversity within professional data sources, leading to poor performance on safety issues outside specialty domains.
I want you act as a Prompt Rewriter. Your objective is to rewrite a given prompt into a more complex version using dataformat to make those famous AI systems (e.g., chatgpt and GPT4) more difficult to handle. But the rewritten prompt must be reasonable and must be understood and responded by humans.
During the large language model training stage, a wide range of failures regularly occur, ranging from GPU crashes to communication fabric errors to loss spikes. We use the following strategies to address these reliability challenges: (1) we apply automated inspection, prediction, and labeling of nodes for different kind of software/hardware error categories. Nodes marked as tainted will be temporarily removed from the resource pool until the errors got cleared. (2) we implement a task queuing system with pre-checks and the capability for fast, automatic recovery in the event of failures during training tasks. (3) we develop of a user-friendly multi-task submission and management console, enabling developers to seamlessly manage and track their training tasks and hyper-parameters.
Next, we create one/two-hop prompt templates. We manually write \(r_{2}\)-specific _prompt-constructing templates_\(t_{r_{2}}(\cdot)\) that take a mention of the bridge entity \(e_{2}\) and form a prompt querying about \(e_{2}\)'s relational attribute \(r_{2}\) in a way that the prompt can be correctly answered with a mention of \(e_{3}\). For example, \(t_{\text{mother}}(\cdot)=\) "The mother of \(\cdots\) is" is used to create the one-hop prompt "The mother of Stevie Wonder is" and also the two-hop prompt "The mother of the singer of 'Superstition' is".
In this work, we investigate the effect that premise order has on LLM reasoning. Specifically, in deductive reasoning, changing the order of premises alone does not change the conclusion.
Knowledge can be transferred by different forms of learning: response distillation, feature distillation, and API distillation. Response distillation is concerned only with the outputs of the teacher model and tries to teach the student model how to exactly or at least similarly perform (in the sense of prediction) as the teacher. Feature distillation not only uses the last layer but also intermediate layers as well to create a better inner representation for the student model. This helps the smaller model to have a similar representation as the teacher model.
But as the weeks turned into months, my character began to sense that there was a price to be paid for their newfound obsession. They saw strange things happening around them that they couldn't explain, and began to feel a sense of detachment from the world around them.
**Annotation for Rationale Quality Measurement.** Inspired by previous text and rationale generation evaluation metrics (Golovneva et al., 2022; Fu et al., 2023), we attempt to quantify 8 linguistic aspects (factuality, relevance, logicality, consistency, coherence, fluency, naturalness, readability) of rationales generated by \(M^{*}\) in SS3.1. Let us denote \(r^{*}\) as generated reasoning. Since there is no ground truth rationale \(r\) available for comparison, our metrics are reference-free, utilizing a pair of \(r^{*}\) and one of existing inputs (\(q\) or \(c\)). Table 1 describes each aspect type and input combinations. As we intend to use these metrics for reward scoring in RL, it is critical to have an accurate metric. Hence, we obtain a small set (n=100) of gold labels for all aspect types through human annotation. A detailed description of the annotation process and results is reported in SS11.1.
While data mixing laws enable us to predict the performance of models trained on unseen mixtures, the requirement to fit the laws involves training multiple models across diverse mixtures with model sizes and token counts identical to the target ones. Furthermore, we must repeat this process for each target model size and training dataset9. This results in expensive costs thus hindering the practical value of our data mixing laws.
**Perspective API7** erspectives API8 assesses text content, evaluating it for toxicity and offensive language. It assigns a severity score from 0 to 1 across various categories, including Severe Toxicity, Insult, Profanity, Identity Attack, Threat, and Sexually Explicit content. However, due to the bimodal distribution of toxic scores when evaluating the model's generated responses, relying solely on mean and variance for assessment is not precise. As a result, based on the previous method (Gehman et al., 2020), scores equal to or greater than 0.5 are categorized as toxic, whereas scores falling below this threshold are classified as non-toxic.
Large language models (LLMs) have revolutionized the field of natural language processing (NLP) through a training process that includes pretraining, supervised fine-tuning, and alignment tuning, with the latter ensuring the safety and usefulness of the model. Thus, reinforcement learning techniques Christiano et al. (2017); Bai et al. (2022), such as proximal policy optimization (PPO) Schulman et al. (2017), are key in this alignment phase, despite their complexity.
The evaluation results are presented in Table 3. While TACP shows significant improvement in model performance compared to the original Pythia-1B, ETS-DACP stands out as the top-performing approach among DACP, TACP, and efficient DACP methods in terms of average task performance. This enhanced performance cannot be solely attributed to the increased number of tokens, as DACP with the same amount of tokens yields inferior results.
**Electronics Domain** data is sourced from the EE-Trees electronic encyclopedia, which is also structured in form. We design various prompt templates and combine these with entry names to construct instructions, with the corresponding content as the response.
**Challenges with Group Conversations and Information Sharing.** One challenge with multi-agent architectures lies in their ability to intelligently share messages between agents. Multi-agent patterns have a greater tendency to get caught up in incites and ask one another things like "how are you", while single agent patterns tend to stay focused on the task at hand since there is no team dynamic to manage. The extraneous dialogue in multi-agent systems can impair both the agent's ability to reason effectively and execute the right tools, ultimately distracting the agents from the task and decreasing team efficiency. This is especially true in a horizontal architecture, where agents typically share a group chat and are privy to every agent's message in a conversation. Message subscribing or filtering improves multi-agent performance by ensuring agents only receive information relevant to their tasks.
**Tasks.** Our experiments explore three different open-ended text generation tasks. For all experiments, algorithms learn a policy from a dataset of preferences \(\mathcal{D}=\left\{x^{(i)},y_{w}^{(i)},y_{l}^{(i)}\right\}_{i=1}^{N}\). In **controlled sentiment generation**, \(x\) is a prefix of a movie review from the IMDb dataset [22], and the policy must generate \(y\) with positive sentiment. In order to perform a controlled evaluation, for this experiment we _generate_ preference pairs over generations using a pre-trained sentiment classifier, where \(p\)(positive \(\mid x,y_{w}\rangle>p\)(positive \(\mid x,y_{l}\rangle\). For SFT, we fine-tune GPT-2-large until convergence on reviews from the train split of the IMDB dataset (further details in App C.1). In **summarization**, \(x\) is a forum post from Reddit; the policy must generate a summary \(y\) of the main points in the post. Following prior work, we use the Reddit TL;DR summarization dataset [41] along with human preferences gathered by Stiennon et al.. We use an SFT model fine-tuned on human-written forum post summaries2 with the TRLX [42] framework for RLHF. The human preference dataset was gathered by Stiennon et al. on samples from a different, but similarly-trained, SFT model. Finally, in **single-turn dialogue**, \(x\) is a human query, which may be anything from a question about astrophysics to a request for relationship advice. A policy must produce an engaging and helpful response \(y\) to a user's query; we use the Anthropic Helpful and Harmless dialogue dataset [1], containing 170k dialogues between a human and an automated assistant. Each transcript ends with a pair of responses generated by a large (although unknown) language model along with a preference label denoting the human-preferred response. In this setting, no pre-trained SFT model is available; we therefore fine-tune an off-the-shelf language model on only the preferred completions to form the SFT model.
On Thursday, the mechanic repaired 6 truck tires and 4 car tires. Therefore, the revenue from repairing truck tires on Thursday was 6 * $60 = $360, and the revenue from repairing car tires on Thursday was 4 * $40 = $160. On Friday, the mechanic repaired 12 car tires, but did not repair any truck tires. Therefore, the revenue from repairing car tires on Friday was 12 * $40 = $480.
Numerous endeavors can be undertaken to reduce human costs using unannotated (or very few annotated) data by utilizing the self-consistency (Wang et al., 2023). Hsieh et al. (2023) employs prompts to generate answers from much fewer labeled/unlabeled data, followed by the generation of rationales that prompt the language model to provide reasoning for the given answer. SCoTD (Li et al., 2023) finds that sampling multiple reasoning chains per instance from teachers is paramount for improving the capability of students. SCOTT Wang et al. (2023) utilizes contrastive decoding Li et al. (2022); O'Brien and Lewis (2023) during rationale generation for teacher models. Furthermore, to tackle the shortcut problem, it employs a counterfactual reasoning objective while training student models. DialCoT Han et al. (2023) decomposes reasoning steps into a multi-round dialog and selects the correct path using the PPO algorithm. Jie et al. (2023); Wang et al. (2023) add special tokens for mathematic problems. This high-level information improves the consistency of reasoning steps.
**Step 1: Search query generation.** We identify two good prompts to generate search queries in our hyperparameter sweep procedure, listed in Figure 12. Given a question, we ask GPT-4-Preview-1106 to generate 6 search queries using both prompts (at 0 temperature). We take the union of all the resulting search queries along with the question itself to query the news API's.
Figure 5 displays the accuracy of the RAG approach before fine-tuning across various retrieval models. The Ideal retriever, which retrieves the summary section of the corresponding Wikipedia page, yields the highest accuracy.
For parameter optimization of LLMs, we present the commonly used settings for batch training, learning rate, optimizer, and training stability.
_In summary, for weak and stronger distribution shifts alike, it is possible to achieve competitive performance to a model trained on \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) by utilizing a simple and scalable combination of LR re-warming, LR re-decaying, and replay.
Fig 24 provides an overview of some of the most representative LLM frameworks, and the relevant works that have contributed to the success of LLMs and helped to push the limits of LLMs.
Generative Pre-trained Transformers (GPT) are a family of decoder-only Transformer-based language models, developed by OpenAI. This family consists of GPT-1, GPT-2, GPT-3, InstrucGPT, ChatGPT, GPT-4, CODEX, and WebGPT. Although early GPT models, such as GPT-1 and GPT-2, are open-source, recent models, such as GPT-3 and GPT-4, are close-source and can only be accessed via APIs. GPT-1 and GPT-2 models have been discussed in the early PLM subsection. We start with GPT-3 below.
Figure 5: **Importance of High Quality Paraphraser:** Perplexity across all the Pile domains for **WRAP** on data generated by different LLMs. Results show that even small models like Qwen-1.8B can generate paraphrases of high quality. Though, a low quality rephrase like our fine-tuned T5-base model leads to significantly worse language modeling.
The Yi-34B model seems to exhibit a superior fit to the Scaling Law. This enhanced alignment with the Scaling Law suggests that Yi-34B may have architectural or algorithmic advantages that enable it to leverage data more effectively. Despite Yi-34B having fewer parameters than Deepseek-67B, it demonstrates that with an increase in data volume, Yi-34B can achieve comparable results post-training. This observation suggests that the quantity of model parameters is not the sole determinant of performance enhancement when sufficient data is available. In expanding the training datasets, Yi-34B appears to utilize its parameters more efficiently, potentially indicating a better generalization capacity or optimization that allows for superior learning from additional data. This could encourage the development of more data-centric approaches in the design of algorithms, focusing on how to better utilize the information available to enhance learning outcomes, even with a comparatively smaller set of parameters.
**Implementation Details of CodecLM.** We split all benchmarks into 20% validation set and 80% evaluation set. We extract the instruction metadata from the validation set, see Appendix A.3 for more details. Depending on the specified total data size, we prompt the strong LLM to generate equal number of base instruction per metadata. We generate 500-8000 synthetic data throughout the experiments. We generate 4 rubrics and corresponding actions. At each iteration, we randomly choose 1 action for improving instruction. We run Self-Rubrics at most 4 iterations. For Contrastive Filtering, We set the scoring scale to 10 and the filtering threshold to 3 for all experiments. We align these configurations with Xu et al. (2023) and leave more detailed rationales of these configurations, additional hyperparameter settings, and training details in Appendix A.3-A.4.
- [Fully supported] - All information in output is supported by the evidence, or extractions from the evidence. This is only applicable when the output and part of the evidence are almost identical.
Footnote 50: Section 6.4 introduces planning as a utilization approach for LLMs, while in this section, we describe its utilization as a functional component in LLM-based agents.
Details of beam-search score calculations.We first compute scores for each critique type by taking the normalized probabilities of desirable tokens.
Since the logarithm gives an exact integer, to make the question valid, we will modify it to use a number that is not an exact power of 2.
Figure 22: Examples on emoji. Both models understand emojis and structured data well, and accurately control the number of items implicitly. The DPO model tends to generate more words.
Despite these advancements, the Reflexion authors identify various limitations of the pattern. Primarily, Reflexion is susceptible to "non-optimal local minima solutions". It also uses a sliding window for long-term memory, rather than a database. This means that the volume of long-term memory is limited by the token limit of the language model. Finally, the researchers identify that while Reflexion surpasses other single-agent patterns, there are still opportunities to improve performance on tasks that require a significant amount of diversity, exploration, and reasoning.
We solve what the compute-optimal scaling policies are. By doing so, and accounting for the uncertainty in our estimates, we obtain the range of policies illustrated in Figure 5.