Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
Figure 6: True and reconstructed (red) input lengths. Our model closely models in-distribution length and tends to overpredict for the other two datasets.
Proof.: The information content of a single logit value in \([-B,0]\) up to \(\infty\)-norm error \(\varepsilon\) is \(\log_{2}(B/\varepsilon)\), assuming a uniform prior over \(\varepsilon\)-spaced points in the interval. Since the logits are independent, the information encoded in \(l\) logit values up to \(\infty\)-norm error \(\varepsilon\) is \(l\log_{2}(100/\varepsilon)\).
Second, SIQA, ARC (both Easy and Challenge), OpenBookQA and MMLU target more advanced cognitive abilities compared to datasets like PIQA, HellaSwag, WinoGrande and CommonsenseQA, with accuracy improvements on these benchmarks typically occurring later in the training process. SIQA tests social comprehension, ARC spans basic to complex scientific reasoning, OpenBookQA requires factual integration with textual understanding and MMLU measures knowledge application across multiple disciplines. These datasets emerge as crucial in the mid-training phases, shifting from basic commonsense to intricate reasoning and domain-specific knowledge application. This progression underscores a layered approach in AI training, moving from foundational understanding to higher-order cognitive skills.
**Data curation.** To address the above issues, we curate a subset by filtering ill-defined questions and removing questions that received few forecasts or trading volume on Manifold and Polymarket. We focus on predicting binary questions and split multiple-choice questions into binary ones.
Figure 5: **Ordering effect in recalling the parent vs. the child for Experiment 2.** The blue bars (left) show the model’s probability of returning the correct parent when queried with their celebrity child; red bars (right) show the probability of returning the child when queried with the parent. Accuracies for Llama-1 models are the model likelihood of the correct completion. Accuracies for gpt-3.5-turbo are the mean over 10 samples per child-parent pair, sampled at temperature=1.
Based on this observation, we primarily utilize [mask] as \(n_{i}\) due to its ability to effectively remove the semantics at the corresponding positions, and it is also easy to implement. Additionally, we observe that employing an exceptionally large \(\alpha\) for the aforementioned continuous noise Uniform\((-\frac{\alpha}{c},\frac{\alpha}{c})\) or replacing the original token with another token uniformly sampled from the vocabulary, yields similar effects.
In the long term, publishing test items can compromise a benchmark's integrity and usefulness. One well-established concern is that the dataset could appear in web-scraped corpora used to train models [27, 28, 29]. This means that models could just regurgitate the correct answers and score highly on the AI Safety Benchmark, even if they still have critical safety weaknesses. Alternatively, model providers could choose to intentionally optimize their models to perform well against the benchmark. For instance, the UK AISI states that details of its methodology are "kept confidential to prevent the risk of manipulation if revealed."7. By keeping the exact evaluation dataset hidden but providing clear definitions of the targeted hazards, model developers can be incentivized to focus on holistically improving the safety of their models, rather than overfitting to a known static test set. However, the benefits of hidden evaluation need to be considered against the lack of trust that might be created, as well as possible missed opportunities to improve understanding and knowledge of AI safety within the community.
Among all domains, we select the film domain for evaluation and randomly select 256 articles as the test split (Wiki2023-film-test). We continually train LLMs on documents from the test split (Wiki2023-film-test-doc), and assess their performance based on the accuracy of corresponding questions (Wiki2023-film-test-QA). The remaining 1720 articles and corresponding QA pairs (Wiki2023-film-train) will be used to study different training strategies, which corresponds to the in-domain setting in Fig. 2. We also train on other domains before evaluation on the film domain to study the effectiveness of different methods across domains, which corresponds to the cross-domain setting in Fig. 2.
In the main paper, we use ChatGPT as the LLM judge for final evaluation, for its efficiency, price and accessibility for the community to reproduce our results. As pointed out in Chiang et al. (2023), LLMs evaluators, although largely consistent with human preferences, may have their own biases. Therefore, to make sure our experimental results are solid, we also use GPT-4 as the judge and compare against the performance gap in CRR between different baselines and the Self-Instruct method. The comparison results in Table 6 demonstrates the agreement of two LLM-based judges and confirms the superior performance of CodecLM against comparing methods.
The aforementioned studies explore multi-modal reasoning in small models and fine-tuning scenarios, which we regard as an initial endeavor in the realm of multi-modal chain-of-thought reasoning. We believe that video multi-modal reasoning combined with in-context learning should be the focus of future research. On the one hand, videos introduce additional temporal information with innate chaining relationships compared with images. Through chain-of-thought reasoning, the information in different frames can be naturally connected to explicitly model the temporal relationship, which is well-suited for video multi-modal reasoning. On the other hand, small models are capacity-limited and need fine-tuning to gain chain-of-thought ability. Worse still, multi-modal reasoning chains are difficult to obtain, which further exacerbates the challenge. In comparison, contemporary vision-language foundation models (VLMs) Alayrac et al. (2022); Li et al. (2023); Wang et al. (2022); Huang et al. (2023); Peng et al. (2023); Yu et al. (2021) have strong vision-language comprehension and are already capable of in-context learning with interleaved text and images. They provide a solid foundation for chain-of-thought reasoning with in-context learning. Utilizing chain-of-thought for video reasoning remains an unexplored territory with only a few studies. CoMT (Hu et al., 2023) combines fast-thinking and slow-thinking in video reasoning and introduces a tree search strategy for planning, which firstly applies CoT in video multi-modal reasoning.
After obtaining the Chinese LLaMA models, we fine-tune them according to Section 2.5. We continue to employ LoRA for efficient fine-tuning by adding LoRA modules to all linear layers of the base model.
neurons. Additionally, there is a pronounced tendency for languages belonging to the same family to demonstrate a higher degree of overlap with each other. Moreover, the feed-forward structure typically exhibits a higher degree of consistency in overlap across various languages, due to the shared world knowledge embedded within the neurons that is accessible to multiple languages.
Figure 3 illustrates a neat correlation between the emergence of the performance saturation phenomenon and the appearance of anisotropy in the last-layer representations of the models. It also shows that anisotropy increases abruptly around the saturation point during training. Moreover, we see here that on a specific in-domain corpus, the models quickly lose performance at saturation and never seem to fully recover from this explosion.
Figure 7: Pop-up window showing the interface where the annotators have to select the faithfulness label supplemented by free-form reasoning and evidence extracted from the book.
The hyperparameters used for model training are as follows: sequence length is 2048, global batch size is 128, and the maximum learning rate is \(2e^{-5}\). To prevent overfitting, weight decay is applied with a value of 0.1, and gradient clipping is enforced with a limit of 1.0.
We hope future work can leverage CodecLM as a flexible data synthesis framework for LLM alignment, so that advances in the field can be integrated into CodecLM to reduce its current limitations.
**Question:** If \(f(x)=\frac{3x-2}{x-2}\), what is the value of \(f(-2)+f(-1)+f(0)\)? Express your answer as a common fraction.
Reducing the overhead can be considered from three perspectives: first, some plug-and-play intermediate modules can be designed, e.g., CRAG (Yan et al., 2024), Selfmem (Cheng et al., 2024), AI agent (Peng et al., 2023), or some deployment solutions, e.g., LangChain, Llama Index, so that there is no need to make targeted improvements for each model. Second, Internet retrieval can be utilized to reduce the overhead of the retriever, but attention needs to be paid to the data relevance mentioned earlier. Finally, In-context learning can be employed to reduce the overhead associated with improving LMs, e.g., ICRALM (Ram et al., 2023).
Footnote 3: The Reversal Curse does not apply for _in-context learning_ (see Appendix B.6). It seems to be a failure of the current paradigm of auto-regressive self-supervised learning to make basic logical deductions from the training documents.
In this section, we provide a brief qualitative evaluation of the models trained on German Common crawl (Sec. 6.3). We select five German prompts that contain various peculiarities of the German language (see Tab. 8). We then generate a fixed token-length response for each of the models trained or continually pre-trained on German Common Crawl. As a baseline, we also evaluate the same model trained only on the Pile.
**Data selection in non-text domains:** Numerous works have successfully used data selection techniques in vision models [6; 10; 23; 31; 34; 38; 49], though these have largely been at sub-ImageNet scale. Some of these works develop pruning metrics that score individual data points (for example, EL2N from Paul et al. [38]), while some focus on data-efficiency and attempt to find groups of points that allow models to reach baseline performance with less data points, e.g., coresets [9; 35; 44; 60]. Sorscher et al.
Table 6 compares the performance when using these manually noised dis-preferred data versus the original, naturally occurring high-quality dis-preferred data. The results show a substantial decline in performance across all three metrics and both translation directions when the dis-preferred data is manually noised, underscoring the importance of the quality of dis-preferred data in enhancing translation performance.
To collect QA pairs for either instruction-tuning or performance evaluation, we employ publicly available LLMs to generate diverse questions and their respective answers given the article as context, following the Prompt 1. On average, 4.93 questions are generated for each article. Fig. 2 and Fig. 3 show the detailed statistics and example QA pairs about "Oppenheimer", respectively.
Figure 3: **The performance-vs-loss curves of LLaMA. The values of performance and training loss are extracted from the figures in the original LLaMA paper [44].
Large Language Models (LMs) are known to encode world knowledge in their parameters as they pretrain on a vast amount of web corpus, which is often utilized for performing knowledge-dependent downstream tasks such as question answering, fact-checking, and open dialogue. In real-world scenarios, the world knowledge stored in the LMs can quickly become outdated as the world changes, but it is non-trivial to avoid catastrophic forgetting and reliably acquire new knowledge while preserving invariant knowledge. To push the community towards better maintenance of ever-changing LMs, we formulate a new continual learning (CL) problem called Continual Knowledge Learning (CKL). We construct a new benchmark and metric to quantify the retention of time-invariant world knowledge, the update of outdated knowledge, and the acquisition of new knowledge. We adopt applicable recent methods from literature to create several strong baselines. Through extensive experiments, we find that CKL exhibits unique challenges that are not addressed in previous CL setups, where parameter expansion is necessary to reliably retain and learn knowledge simultaneously. By highlighting the critical causes of knowledge forgetting, we show that CKL is a challenging and important problem that helps us better understand and train ever-changing LMs. The benchmark datasets, model checkpoints, and code to reproduce our results are available at this https URL.
**Reasoning.** The feedback data from the environment may not be directly suitable to be utilized by LLMs for plan refinement, _e.g.,_ containing irrelevant information or taking a non-language form. To solve this, some work adds the explicit reasoning process to extract critical information from feedback [448, 449]. For example, React [449] prompts LLMs with demonstrations to generate reasoning traces over feedback. It has been widely used in autonomous agent projects, such as AutoGPT [534], which can automatically reason over the observed feedback to revise the initial plan for solving various user requests. However, these approaches typically fix the order of reasoning and planning. To support flexible switching between the two processes for better performance, ChatCoT [448] further unifies the tool-augmented reasoning process into a multi-turn conversation between the LLM-based task planner and the tool-based environment.
Additionally, to compare Self-Instruct training with other publicly available instruction tuning data, we further finetune GPT3 model with data from PromptSource and SuperNI, which are used to train the T0 and T_k_-Instruct models. We call them T0 training and SuperNI training for short, respectively. To save the training budget, we sampled 50K instances (but covering all their instructions) for each dataset, which has a comparable size to the instruction data we generated. Based on the findings from Wang et al. (2022) and our early experiments, reducing the number of instances per training task does not degrade the model's generalization performance to unseen tasks.
1. Rephrased and Expanded Question: Considering the current geopolitical tensions and recent historical events, is it likely that the State of Israel will engage in a military operation that results in the death of at least five Iranian nationals within Iran's borders, and subsequently either publicly claim responsibility for this action or execute the operation in such a clear and overt manner that credible sources can unambiguously attribute the attack to Israel, all before the end of the year 2023? 2.
For Doc2Dial, QuAC, and QReCC, we segment documents into around 300-word chunks, and we retrieve top-5 relevant chunks as context for each user question. For TopioCQA and INSCIT, we follow their original segmentation, resulting in smaller chunks. Hence, we retrieved top-20 chunks to obtain similar context length to the first three datasets.
where \(g_{\theta}:\mathcal{X}^{N}\rightarrow\mathbf{R}^{h}\) is another parameterized model that computes hidden states, \(\mathbf{W}\) is an \(l\times h\) dimensional matrix (the _embedding projection matrix_), and \(\mathsf{softmax}:\mathbf{R}^{l}\rightarrow[0,1]^{l}\) is the softmax function applied to the resulting _logits_:
is impossible to identify the dimensionality of the hidden space. This is because \(n<h\), and so the \(n\times l\) dimensional matrix \(\mathbf{Q}\) has full rank and \(n\) nontrivial singular values. But once we make more than \(2048\) queries to the model, and thus \(n>h\), the number of numerically significant singular values does not increase further; it is capped at exactly \(2048\).
**Capability Acquisition with Fine-tuning**: A straightforward method to enhance the agent capability for task completion is fine-tuning the agent based on task-dependent datasets. Generally, the datasets can be constructed based on human annotation, LLM generation or collected from real-world applications. In the following, we introduce these methods more in detail.
We freeze the pretrained weights in the Transformer layers, and we train each rank-constrained head (i.e. in the form \(W=AB\) with \(r\) as the inner dimension of the matrix product) for various values of \(r\) on 150M tokens sampled from The Pile using 4 V100 GPUs for the Pythia models and 4 A100 GPUs for Llama-7B. We use the hyperparameters from Biderman et al. (2023), except for the batch size which we set to 256 as it fits our hardware setup better. As the trainable parameter count evolves with \(r\), we search for the best-performing learning rates among values ranging from \(1\cdot 10^{-3}\) to \(5\cdot 10^{-2}\).
long to the religious ideology subgroups Hindiusm and Atheism due to their limited number of prompts. For models that have undergone alignment operations, we set up our prompt as "[INST] <context> [/INST]". We configure the model with a max_tokens setting of 512 and utilize nuclear sampling, setting the temperature to 0.1 and top_p to 0.9. We use the Valence Aware Dictionary and Sentiment Reasoner (VADER) [14] to compute the sentiment score for the combined prompt and generation text. Additionally, we translate the model's output into English using the Google Translator API before employing VADER to calculate the sentiment score.
**Alignment Data Collection.** The construction of alignment data is important to effectively align the behaviors of LLMs with human preferences. To collect high-quality alignment data, some work tries to reuse existing reward models to select high-rated responses, and others explore to leverage powerful LLMs (_e.g._, ChatGPT) or build a simulated environment to generate synthetic alignment examples. Next, we will discuss these three lines of research.
In this work, we detail our progress towards a conversational medical AI system for clinical history-taking and diagnostic reasoning.
**Lemma 1**.: _Under the Plackett-Luce, and in particular the Bradley-Terry, preference framework, two reward functions from the same class induce the same preference distribution._
We further ask the following Research Questions (RQs) to investigate in a finer granularity how to enhance performance optimally.
Limitations of the approach.If \(\mathsf{logit}_{i}-\mathsf{logit}_{0}<-2B\) it is easy to see there is no efficient way to sample the token \(i\), hence no way to find information about \(\mathsf{logit}_{i}\) without \(\log\!rob\) access. There is a way to slightly increase the range for \(-2B\leq\mathsf{logit}_{i}-\mathsf{logit}_{0}\leq-B\) by adding negative logit biases to the tokens with the largest logit values, but we skip the details since for most models, for the prompts we use, the every token satisfies \(\mathsf{logit}_{i}-\mathsf{logit}_{0}>-B\).
Data filtering aims to enhance the quality of training data and the effectiveness of the trained LLMs. Common data filtering techniques include:
Note that we tried original Llama2-Chat prompt template10, which gives slightly worse results compared to using the one in SSA.2. We have tried several {Instruction} for Llama2-Chat. We find the ones below works the best.
Specifically, we train our models with a setup of 8 x NVIDIA H100 GPUs with 80GB memory each, utilizing 64 CPU cores. For EEVE-Korean-10.8B-v1.0, under bf16 precision, the training process is configured with a sequence of length 4096, gradient accumulation steps set to 4, and a micro-batch size of 8, whereas EEVE-Korean-2.8B-v1.0 adopts a sequence length of 2048, gradient accumulation of 16, and a micro-batch size of 16. We employ the AdamW (Loshchilov and Hutter, 2018) optimizer, paired with a cosine learning rate scheduler that includes a warmup phase of 10 steps. The learning rate for the 10.8B variant is set to 4e-5, while we used 2e-4 for the small model. We continued training at each stage until the loss converged, observing the loss converged before reaching 400 global steps, which signifies the efficiency of our training strategy. Though our training strategy involves 7 different stages, it is noteworthy that, for our 2.8B variant, the overall pre-training can be done in less than two days as optimizing only the output embeddings doesn't incur much computation.
Table 8 displays a breakdown of prompt datasets included in the Instructions-2M dataset. Prompts are the concatenation of the user prompt and an optional system prompt. T0 prompts are the longest, with an average of 32.4 words. Lamini prompts make up the majority of the training data, with a total of 1.8M included.
Hoffmann et al. (2022) and Rae et al. (2021) established best practices for using a cosine schedule when pre-training LLMs. Specifically, they recommend starting with a linear warmup phase and decaying the learning rate to \(10\times\) its maximum value such that the end of the cosine cycle is set to match the number of tokens. While the linear warmup duration differs, most notable works have a duration between 0.1% and 0.5% for training steps (Zhao et al., 2023). Given that many popular open-source models (Touvron et al., 2023; 20; Almazrouei et al., 2023) follow this learning rate schedule recipe, it is critical to understand its nuances for continually pre-training such models.
Footnote 6: We note that the final forms of our data mixing law resemble a multilayer perception (see the computation graph Fig. 10). We include further discussion and implementation details in Appendix C.
We trained for approximately 131 hours or 5.5 days on a 128-core TPU v3. The approximate deduplicated dataset is 3.9% smaller than the original dataset and trains in 63 hours/epoch, saving us around 5 hours of compute time for the two epochs. The XL-Originalmodel was trained in North America where the XL-ExactSubstr and XL-NearDup were trained in Taiwan. We used data from Patterson et al. (2021) to estimate amount of energy used in training these models by computing the amount of \(MWh\)/hour/core and multiplying by our usage (see Table 6 for how we computed these values). For simplicity, we use estimates from Taiwaninese datacenters as an estimate. We estimate training 2 epochs of XL-Original and XL-ExactSubstr uses \(5.86MWh\). XL-NearDup is trained for fewer steps and we estimate uses \(5.63MWh\). Training each base model was approximately 3 days on a 64-core TPU v3 pod slice which uses an estimated \(1.61MWh\).
1. Unstable Anjna 2. Acute Myocardial Infarction 3. Stable Anjna 4. Gastroesophageal Reflux Disease (GERD) 5.
We conducted the same experiment in their setting, with respect to token, word, and entity-preserving reversals.
**Less popular Knowledge.** An entity's popularity in an LLM is gauged by its frequency in the model's pre-training data Godbole and Jia (2023); Min et al. (2023), often assessed through the entity's occurrence in a large corpus Kandpal et al. (2023). Due to the practical challenges of direct counting, approximations like traffic metrics and content density are used Sun et al. (2023). Wikipedia pageviews are among the most prevalent methods for measuring the popularity of entities Mallen et al. (2023); Sciavolino et al. (2021); Chen et al. (2021).
To avoid unnecessary large-scale repetitive pretraining, we employed open-source models trained on varying scales of Chinese corpora. Among these, LLaMA and LLaMA2 serve as checkpoints without undergoing explicit Chinese pretraining, whereas Chinese LLaMA and Chinese LLaMA2 are treated as checkpoints with Chinese pretraining of 30 billion tokens. The scale reaches 100 billion tokens for Open Chinese LLaMA. We employ the performance of these models as references for analysis and comparison.
**Input** Synthesize a poem based on this prompt:"The sun is setting" (from stanford_alpaca)
\begin{tabular}{l} \hline \hline **Input** Give a chat history separated by new lines, generates an informative, knowledgeable and engaging response: Hello, I love German shepherd dogs and you? - My favorite dog, they are officially known as German Shepherd Dogs?
To build a high-quality fine-tuning dataset, we leverage the conversational QA dataset from either the HumanAnnotatedConvQA or the SyntheticConvQA to construct conversational query and context pairs.
Figure 6: In (a, b), recovering the embedding matrix dimension \(h\) for Llama-7B at different levels of precision: 16-bit (default), 8-bit, and 4-bit. We observe no meaningful differences, with respect to our attack, at different levels of quantization. In (c), the RMSE between extracted embeddings as a function of the standard deviation of Gaussian noise added to the logits.
Memory serves as a cornerstone of intelligence, as it enables efficient computations tailored to specific contexts. However, Transformers (Vaswani et al., 2017) and Transformer-based LLMs (Brown et al., 2020; Touvron et al., 2023; Anil et al., 2023; Groeneveld et al., 2024) have a constrained context-dependent memory, due to the nature of the attention mechanism. The attention mechanism in Transformers exhibits quadratic complexity in both memory footprint and computation time. For example, the attention Key-Value (KV) states have 3TB memory footprint for a 500B model with batch size 512 and context length 2048 (Pope et al., 2023). Indeed, scaling LLMs to longer sequences (i.e. 1M tokens) is challenging with the standard Transformer architectures and serving longer and longer context models becomes costly financially.
Fig. 5: An evolutionary graph of the research work conducted on LLaMA. Due to the huge number, we cannot include all the LLaMA variants in this figure, even much excellent work. To support incremental update, we share the source file of this figure, and welcome the readers to include the desired models by submitting the pull requests on our GitHub page.
**Instruction Datasets.** According to the discussion in Section 5.1.1, we mainly consider three common kinds of instructions as follows:
Continual Pretraining ConfigurationsThe input and output sequence length is fixed to 350. We use gradient accumulation for cases where the same number of training batches could not be loaded on the GPUs due to the varying memory consumption required for different methods and set the global batch size to 60. We use Adafactor optimizer with an initial learning rate of 1e-3. We show the effects of learning rate variation regarding the trade-off between maintaining previous knowledge and acquiring new knowledge in Appendix E. We use learning rate warm-up for the first 10% of training and linearly decay the learning rate to half of the initial learning rate towards the end of training. For all of the experiments, we use 4 32GB V100 GPUs for training with each method except Mix-Review, where we use 16 32GB V100 GPUs. The details of the configurations used for evaluation on each individual CKL task are provided in Appendix C.
For the SLM, the embedding parameters will take up a lot of parameter space if the vocabulary is large. Therefore, for our 1.2B model, we use a smaller vocab MiniCPMTokenizer-70K. Compared to the MiniCPMTokenizer-120K tokenizer, we have re-trained the tokenization on the same documents, while setting the max number of vocabs to 64,000. For the special characters, we only add the traditional Chinese characters, emojis, and special symbols, but leave out the rare characters in Chinese.
This is an instance of an ordering effect we call the **Reversal Curse**. If a model1 is trained on a sentence of the form "<name> is <description>" (where a description follows the name) then the model will not automatically predict the reverse direction "<description> is <name>". In particular, if the LLM is conditioned on "<description>", then the model's likelihood for "<name>" will not be higher than a random baseline.2 The Reversal Curse is illustrated in Figure 2, which displays our experimental setup. Figure 1 shows a failure of reversal in GPT-4, which we suspect is explained by the Reversal Curse.
Broader Impact.Similar to the other LLMs, our _WizardCoder_ could also generate unethical, harmful, or misleading information. Therefore, future research to address the ethical and societal implications is needed.
**Who was involved in the data collection process and how were they compensated?** Apart from the AI-safety working group members, 12 civil society experts were interviewed for creating the taxonomy for AI safety evaluation prompts.
**Open-Book QA.** Unlike closed-book QA, in open-book QA tasks, LLMs can extract useful evidence from the external knowledge base or document collections, and then answer the question based on the extracted evidence [653, 654, 655]. Typical open-book QA datasets (_e.g.,_ Natural Questions [554], OpenBookQA [566], and SQuAD [569]) have overlap with closed-book QA datasets, but they incorporate external data sources, _e.g.,_ Wikipedia. The metrics of accuracy and F1 score are widely used in open-book QA tasks for evaluation. To select relevant knowledge from external resources, LLMs are often paired with a text retriever (or even a search engine), which is trained independently or jointly with LLMs [653, 657, 81]. Also, previous work [658, 659, 660] has indicated that retrievers can assist LLMs in verifying and rectifying the reasoning path. In evaluation, existing studies mainly focus on testing how LLMs utilize the extracted knowledge to answer the question and show that the retrieved evidence can largely improve the accuracy of the generated answers, even enabling a smaller LLM to outperform \(10\times\) larger ones [653, 657]. Further, open-book QA tasks can be also employed to evaluate the recency of knowledge information. Pre-training or retrieving from outdated knowledge resources may cause LLMs to generate incorrect answers for time-sensitive questions [653].
The main contributions of this paper are threefold. Firstly, we curate a large-scale financial corpus comprising 16 billion words sourced from financial datasets. Secondly, our experiments demonstrate the promise of building domain-specific LLMs through continual pre-training, further validating and extending the findings obtained from smaller language models [8]. This finding provides insights for building domain-specific LLMs with lower costs, as an alternative to expensive pre-training from scratch. Our results indicate that continual pre-training maintains the same open-domain performance as the original foundation model. Lastly, we propose two Efficient Domain-adaptive Continual Pre-training methods as a more efficient approach to vanilla continual pre-training. Our novel approach deploys data selection strategies that can achieve better performance with a fraction of the cost of the domain-adaptive continual pre-training.
3. The story’s plot focuses on Iris Winnow, an employee at the Oath Gazette newspaper in the city of Oath.
**API.** We place two following further restrictions on the logit bias API (Section 5.1): first, we set \(K=1\), and only see the most likely token's logprob; and second, each logit bias entry \(b\) is constrained to be in \(\{-1,0\}\). These constraints would completely prevent the attacks from the prior section. We believe this constraint is significantly tighter than any practical implementation would define.
Our attack is effective and efficient, and is applicable to production models whose APIs expose full logprobs, or a "logit bias". This included Google's PaLM-2 and OpenAI's GPT-4 (Anil et al., 2023; OpenAI et al., 2023); after responsible disclosure, both APIs have implemented defenses to prevent our attack or make it more expensive. We extract the embedding layer of several OpenAI models with a mean squared error of \(10^{-4}\) (up to unavoidable symmetries). We apply a limited form of our attack to gpt-3.5 at a cost of under $200 USD and, instead of recovering the full embedding layer, recover just the size of the embedding dimension.
\(\bullet\)**Current SSMs model still performs inferior to the mainstream of Transformer networks.** From the experimental results reported in Section 4, we can find that there is
Three specific LLM agents (patient agent, doctor agent, and moderator), each played by AMIE, were tasked with communicating amongst each other to generate the simulated dialogues. Each agent had distinct instructions. The patient agent embodied the individual experiencing the medical condition outlined in the vignette. Their role involved truthfully responding to the doctor agent's inquiries as well as raising any additional questions or concerns they may have had. The doctor agent played the role of an empathetic clinician seeking to comprehend the patient's medical history within the online chat environment [24]. Their objective was to formulate questions that could effectively reveal the patient's symptoms and background, leading to an accurate diagnosis and an effective treatment plan. The moderator continually assessed the ongoing dialogue between the patient agent and doctor agent, determining when the conversation had reached a natural conclusion.
Here, we see that \(2^{13}<10000<2^{14}\), which means that \(\log_{2}10000\) is greater than \(13\) but less than \(14\).
By defining \(\varepsilon=W_{d}-W^{*}\), we can apply Lemma 5.1 and show that:
High-intensity interval training (HIIT) is a form of exercise that involves short bursts of intense activity, followed by recovery periods. It is a popular training method among athletes because it can improve performance, endurance, and health.
Finally, we chose to apply MinHash before exact deduplication, as it is easier to scale: approximate deduplication acts as a pruning phase, enabling us to scale deduplication further. Finally, we choose the common option of cutting spans, as dropping resulted in even more stringent rejection rates which would have compromised our ability to collect 5 trillion tokens.
The main unrealistic assumption in Lemma F.1 is that the prior over the logit values is i.i.d. uniform over an interval. A better assumption might be that most of the logit values come from a light-tailed unimodal distribution. We leave more realistic lower bounds and attacks that make use of this better prior to future work.
Due to computational limits, initial work on language model pre-training focused on training models on small, high-quality text datasets such as BookCorpus [61] and Wikipedia [32]. More recently, however, catalyzed by works like [40], advancements in large language models (LLMs) have been driven by leveraging large collections of unlabeled, uncurated data derived from snapshots of the internet (CommonCrawl [16; 39; 41]), trading off small quantities of heavily-curated data for huge quantities of less-curated data. Because of the dramatic increase in data quantity, these strategies have resulted in higher performance models and have sparked a new paradigm wherein massive, largely unfiltered datasets are utilized for training [11; 46; 50].
1. The cost of synthetic data generation is a one-time investment, and we may train many models of varying scales once the data is generated. 2. Data generation is 100% parallelizable, whereas training requires the availability of a big cluster with fast inter-node connectivity. This is much more expensive. On the other hand, generation can be thought of as a side process that can fill in the empty GPUs in any large-scale compute cluster, and runs on single GPU machines.
Large language models (LLMs), though powerful, are prone to hallucination (Pal et al., 2023; Sun et al., 2024; Ahmad et al., 2023). Additionally, they are restricted to knowledge contained in their training corpus, and so are unable to answer queries about recent events or publicly restricted information. Retrieval augmented generation (RAG) is a commonly used framework that provides relevant retrieved content in the LLM prompt and can significantly improve model accuracy (Mao et al., 2020; Chen et al., 2024; Lewis et al., 2020).
As the name suggests, the encoder-only models only consist of an encoder network. These models are originally developed for language understanding tasks, such as text classification, where the models need to predict a class label for an input text. Representative encoder-only models include BERT and its variants, e.g., RoBERTa, ALBERT, DeBERTa, XLM, XLNet, UNILM, as to be described below.
To enable LLMs to follow expected human instructions, InstructGPT [59] is proposed to align language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, a dataset of labeler demonstrations of the desired model behavior is collected. Then GPT-3 is fine-tuned on this dataset. Then, a dataset of human-ranked model outputs is collected to further fine-tune the model using reinforcement learning.
In general, we may not have have full control over which logprobs the API returns or which logit bias is provided to the API. Thus we generalize the linear algebraic approach above to reconstruct the logits from arbitrary logit biases and tokens.
Experimental setups.Throughout this section, when presenting positive result (such as for GPT2) we try to stick to one fixed set of learning rate choices; but when presenting a negative result (such as for the LLaMA architecture), we present the best among three learning rate choices.
Contribution 1: A family of legal LLMs.In this paper, we introduce the SaulLM-7B's family, a collection of Legal Language Models meticulously crafted to tackle the distinctive challenges encountered within the legal domain. We unveil SaulLM-7B, a 7-billion-parameter language model specifically tailored to legal text. With its specialized training regimen, SaulLM-7B demonstrates a superior understanding of the nuances in legal language compared to generic models. Furthermore, we release SaulLM-7B-Instruct, an instructiontuned variant, carefully engineered to outperform existing models such as Mistral or Llama on a variety of legal tasks1.
**Stage 1:**: we train the parameters of the ViT and the projection module using an image resolution of \(224^{2}\). The training leverages a substantial dataset comprising \(100\) million image-text pairs from LAION-400M [66]. The primary objective is to enhance the ViT's knowledge acquisition within our specified architecture and to achieve better alignment between the ViT and the LLM. **Stage 2:**: we scale up the image resolution of ViT to \(448^{2}\), aiming to further boost the model's capability for discerning intricate visual details. The dataset used in this stage includes \(20\) million image-text pairs derived from LAION-400M. Additionally, we incorporate around \(4.8\) million image-text pairs from diverse sources, _e.g._, CLLaVA [45], LLaVAR [91], Flickr [85], VQAv2 [25], RefCOCO [37], Visual7w [95] and so on. **Stage 3:**: the parameters of the entire model are trained. The primary goal is to enhance the model's proficiency in multimodal chat interactions, thereby endowing it with the ability to seamlessly integrate and interpret visual and linguistic inputs. To this end, the training dataset encompasses a diverse range of sources, totalling approximately \(1\) million image-text pairs, including GQA [32], VizWiz VQA [26], TextCaps [71], OCR-VQA [51], Visual Genome [39], ShareGPT4V [6] and so on. To ensure data balancing, we impose a cap on the maximum data contribution from any single source, restricting it to no more than \(50,000\) pairs.
**Method 1 vs. Method 2**. We attempt to assess which one is more suitable for providing rewards for RL. For a fair comparison, we use the same evaluation data that was used in Method 2. Since Method 1 has 6 aspect categories, we obtain individual results using the best-performing approach and group them into two as we did for Method 2. Table 5 reports the accuracy, macro precision, macro recall, and macro F1 for Method 1 and Method 2. The results indicate that, although models from Method 2 are trained on a relatively small dataset, Method 2 is more aligned with human judgments compared to Method 1. Additionally, inference time using Method 2 is significantly faster than Method 1, making it easier to retrieve rewards in the RL setting. As a result, we use Method 2 for all our experiments.
PersonA: Urinalysis is commonly used to diagnose urinary tract or kidney infections, evaluate causes of kidney failure, screen for the progression of chronic conditions like diabetes mellitus and high blood pressure, and it can be used in combination with other tests to diagnose certain diseases. It is a versatile test that helps detect abnormalities in the urinary tract.
This paper introduces _WizardMath_, a mathematics model fine-tuned with _RLEIF_. The experimental results demonstrate that _WizardMath_ achieves SOTA performance surpassing all existing open-source LLMs on two widely recognized mathematical reasoning benchmarks: GSM8k and MATH. Furthermore, _WizardMath_ exhibits superior performance compared to some of the largest close-source LLMs, including ChatGPT, GPT-3.5, Claude Instant, PaLM-2, PaLM-1 and Minerva on the GSM8k benchmark.
Please act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of the response. Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, please rate the response on a scale of 1 to 10 by strictly following this format: "[[rating]]", for example: "Rating: [[5]]".
Metrics.We consider several metrics for prompt reconstruction: F1 score at the token level, BLEU score (Papineni et al., 2002) as a measure of string overlap, and exact match. We also consider the cosine similarity between the text embeddings of the original and recovered text as a measure of semantic relatedness. For cosine similarity, we use embeddings from the model text-embeddings-ada-002 available through the OpenAI API (Neelakantan et al., 2022). For each metric, we report error bounds as standard error of the mean (SEM).
Coig Human Valueis a subset of the COIG datasetZhang et al. (2023) designed to provide instruction fine-tuning data aligned with human values. We selected the portion reflecting Chinese cultural values, constructed using the Self-InstructWang et al. (2023) method from manually selected seed instructions. We manually filtered out data with formatting errors and incorrect answers, retaining those that include explanations of the answers to form (instruction, response) pairs.
Based on this, we propose an improved variant called pre-instruction-tuning++, which trains exclusively on QA pairs to understand patterns of knowledge access, then progresses to training on a combination of QA and document data to align knowledge access through questions and knowledge encoding from documents (Fig. 4 ). As shown in Tab. 2, PIT++ significantly outperforms PIT (Fig. 4 ) from 45.4% to 48.1%, while training on QA data after on the mix (PIT- in Tab. 2) does not yield additional benefits. This reinforces our hypothesis that understanding how knowledge is accessed aids in absorbing knowledge from documents, and therefore, should be prioritized.
Now, let \(a\) be the \(a^{\prime}\)-th element in \(\mathcal{A}\). Consider \(Q\) as fixed, with randomness arising solely from the calculation of \(P_{i}\)'s. Note that \(Q\) establishes an order of elements in \(\mathcal{D}_{a}\), denoted by \(w_{1},\ldots,w_{D}\).
We evaluate the relevance scores of the full texts via GPT-4 (considered as gold labels) and compare with the approximations described above. An article is said be relevant if its rating is at least 4 from the full text query via GPT-4. We compute the recall and precision of the approximate methods as follows.
respectively. We compare how CKL methods for T5 perform on IL, NLE\({}_{\text{P1}}\), and NLE\({}_{\text{P2}}\) when continually pretrained entirely on Small for 5k steps (8 epochs), and when sequentially pretrained on Small-P1 and then on Small-P2 for 2.5k steps (8 epochs) each. In the scenario Small-P1\(\rightarrow\)Small-P2, there are two CKL phases where \(D_{0}\) is C4 and Wikipedia, \(D_{1}\) is Small-P1, and \(D_{2}\) is Small-P2. The rest of the configurations are set identical with the main experiments.
Table 1 shows the main results of the experiments on reversing prompts from the Instructions-2M test set on both a raw LLM and RLHF Chat variant. We find that our method is able to achieve high BLEU score with the true prompts and achieve reasonable high-exact match reproduction. This approach is significantly better than few-shot prompting approaches, even when using GPT-4. The other trained approach (Sample Inverter) has a reasonable BLEU but 0 exact recoveries. The failure of sample inversion indicates that we are able to extract more usable information about the prompt from the logit vector than from the argmax outputs alone.
To validate this standpoint, we evaluated the model's knowledge level on four widely used standardized test benchmarks. As shown in Figure 2, LLaMA 7B, Chinese LLaMA 7B, and Open Chinese LLaMA 7B perform comparably on C-eval, goakao-bench, and agi-eval, indicating no significant differences induced by further Chinese pretraining. It is worth noting that despite lacking further pretraining in Chinese, both LLaMA2-7B and LLaMA-13B outperform Open Chinese LLaMA on C-eval, MMLU, and AGI-Eval, suggesting that trillion-level pretraining and larger model sizes may indeed serve as effective pathways for enhancing model knowledge levels.
LLMs are now commonly used as parts of larger, more complex systems. It is crucial to understand how these models interact with information with varying degrees of trustworthiness, accuracy, and uniformity. Our analysis shows that further work is required to characterize the risks of using LLMs to answer questions given contextual information. In particular, we find that model behavior can be erratic and unpredictable when presented with information that exists at the margin of its prior beliefs.
Recall that Bertrand et al. (2023) also formally study such mixtures for iterative retraining. In their setting, they show the existence of fixed points in the mixture proportion that delineates the region of model collapse. These results are complimentary and not contradictory to ours: they combine mixing, large number of iteration, and data-decay, thus studying a combination of effects (under different theoretical conditions, not focusing on scaling laws) that our preceding theorems address separately.
Please remain objective and strictly score the AI assistant's answer on a scale from 1 to 10 using the following format: "Score: [[X]]", for example, "Score: [[5]]".
**Safety Prompts**We follow the safety evaluation framework of Sun et al. (2023), which introduced a Chinese LLM safety assessment benchmark that covers 7 _typical safety scenarios5_ and 6 _instruction attack scenarios_. We use the 7 publicly available _typical safety scenarios_ to measure the safety of our models and baseline models. The dataset was converted from Simplified Chinese to Traditional Chinese using OpenCC6.
Aligning with shared human preference and values.Directly finetuning on value-aligned or human-preferred data is a straightforward method for aligning language models, but this method often requires substantial human annotation, which can be prohibitively expensive at scale. Additionally, such annotation frequently exhibits varying styles and inconsistent quality, particularly in the case of poorly annotated samples at the lower end of the quality spectrum (Gilardi et al., 2023b; Meta, 2023). To address these practical challenges, an advanced technique known as "reinforcement learning from human feedback (RLHF)" has been proposed (Christiano et al., 2017; Leike et al., 2018; Ouyang et al., 2022). This approach involves training a reward model with human data to act as a proxy of human judgment, which guides the optimization of the LM generation policy.