Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
One natural defense to our attacks is to obfuscate the logits by adding noise. This will naturally induce a tradeoff between utility and vulnerability--more noise will result in less useful outputs, but increase extraction difficulty. We empirically measure this tradeoff in Figure 5(c). We consider noise added directly to the logits, that is consistent between different queries of the same prompt. To simulate this, we directly add noise to our recovered logits, and recompute the extracted embedding matrix. For GPT-2, we measure the RMSE between the true embedding matrix and the embedding matrix extracted with a specific noise level; for ada and babbage, we measure the RMSE between the noisy extracted weights and the weights we extracted in the absence of noise. We normalize all embedding matrices (to have \(\ell_{2}\) norm 1) before measuring RMSE.
The readers should be note that these findings about open-source language models are limited to the model sizes. We will continually update this part by including the results of larger versions of these models, and also call for the support of computational resources for more experiments.
Finally, we show that aggregates of our system with the crowd forecasts outperform either one in isolation.
Following the revelation of the scaling law Kaplan et al. (2020), there has been a vigorous pursuit in the field of Large Language Models (LLMs) Hoffmann et al. (2022); Bai et al. (2023); Gemini et al. (2023); Chowdhery et al. (2023); Achiam et al. (2023), encompassing models with up to an astonishing number of parameters in the trillions Fedus et al. (2022). These models have emerged as a pivotal driving force in the evolution of artificial intelligence.
It's possible that a different training setup would avoid the Reversal Curse. We try different setups in an effort to help the model generalize. Nothing helps.
where \(l_{j^{\prime}}\) is the logits of each word and \(t\) is the temperature coefficient. Reducing the temperature \(t\) increases the chance of selecting words with high probabilities while decreases the chances of selecting words with low probabilities. When \(t\) is set to 1, it becomes the default random sampling; when \(t\) is approaching 0, it is equivalent to greedy search. In addition, when \(t\) goes to infinity, it degenerates to uniform sampling.
**Gemini:** In [99], Gemini team introduced a new family of multimodal models, that exhibit promising capabilities across image, audio, video, and text understanding. Gemini family includes three versions: Ultra for highly-complex tasks, Pro for enhanced performance and deployability at scale, and Nano for on-device applications. Gemini architecture is built on top of Transformer decoders, and is trained to support 32k context length (via using efficient attention mechanisms).
_Comparison with BloombergGPT_: results reported on FLARE are not directly comparable with results reported in BloombergGPT [31] on the same tasks, as the data splits used are not public. We could not match the performance of publicly available models like OPT-66B or GPT-NeoX reported by [31], on all four tasks. See the detailed comparison between the results in Appendix A.
**Handling Outliers:** Identifying and handling outliers or anomalies in the data to prevent them from disproportionately influencing the model.
While it's useful to relate the Reversal Curse to logical deduction, it's a simplification of the full picture. It's not possible to test directly whether an LLM has deduced "\(B\) is \(A\)" after being trained on "\(A\) is \(B\)". LLMs are trained to predict what humans would write and not what is true (Lin et al., 2022). So even if an LLM had inferred "\(B\) is \(A\)", it might not "tell us" when prompted. Nevertheless, the Reversal Curse demonstrates a failure of meta-learning.
11. Roman Kitt faces challenges balancing his family’s aspirations and his role at the Oath Gazette.
In this section, we validate the workflow presented in Figure 2 by systematically deactivating language-specific neurons across different layers and components to examine their impact on a range of tasks.
General UnderstandingThe General Understanding category comprises datasets testing broader cognitive skills and language comprehension. **ARC Easy (ARC-E)**Clark et al. (2018) is the less challenging counterpart of ARC-C, featuring questions that require basic reasoning skills. **BoolQ**Clark et al. (2019) includes boolean questions that focus on reading comprehension and general language understanding. **Winogrande (Wino.)**ai (2019) challenges models with common sense reasoning in language, particularly in pronoun disambiguation. **PIQA**Bisk et al. (2020) assesses understanding of physical processes, an essential part of practical common sense. **HellaSwag**Zellers et al. (2019) tests the ability to complete scenarios coherently, demanding both language understanding and common sense. **TruthfulQA**Lin et al. (2021) is centered on generating truthful, accurate answers, thus testing the model's factual correctness. **OpenBookQA (OBQA)**Mihaylov et al. (2018) evaluates the understanding of a broad range of facts and concepts. Finally, **LogiQA-2**Liu et al. (2023) assesses the model's capacity to comprehend and apply logical principles.
Shi et al. (2023a) argue that retrieval can produce documents that not only fail to provide helpful information but can also compromise the quality of the language model output. As a result, many scholars in the field of RALM focus on improving the relevance between the retrieved content and the user's input to enhance the final output's quality.
Question: What was the revenue drop in the first quarter compared to the same period last year? Answer: The revenue dropped 15 percent.
The previous sections assume we have access to the full language model output probability vector. However, many language model platforms limit the information returned from API calls. For example, a well-known service's API only exposes either samples or the top-5 log probabilities, but does not expose all output probabilities for a given input.
In this part, we mainly focus on three basic types of ability evaluation for LLMs, _i.e._, language generation, knowledge utilization, and complex reasoning. It is noted that we do not intend to have complete coverage of all the related tasks, but instead only focus on the most widely discussed or studied tasks for LLMs. Next, we introduce these tasks in detail.
**QUESTION:** Jim spends 2 hours watching TV and then decides to go to bed and reads for half as long. He does this 3 times a week. How many hours does he spend on TV and reading in 4 weeks?
Figure 8: Final loss with trained rank-constrained heads (mimicking \(W^{s}_{d}\)), as a function of the theoretical \(W\)-error for rank \(d\) on the head of the Pythia-1B model.
**Baselines with retrievals.** We evaluate models augmented with retrieval at test time or during training. The first category includes standard RAG baselines, where an LM (Llama2, Alpaca) generates output given the query prepended with the top retrieved documents using the same retriever as in our system. It also includes Llama2-FT, where Llama2 is fine-tuned on all training data we use without the reflection tokens or retrieved passages. We also report the result of retrieval-augmented baselines with LMs trained with private data: Ret-ChatGPT and Ret-Llama2-chat, which deploy the same augmentation technique above, as well as perplexity.ai, an InstructGPT-based production search system.
**Position Embeddings.** Since the self-attention modules in Transformer are permutation equivariant, position embeddings (PE) are employed to inject absolute or relative position information for modeling sequences.
Figure 4 displays the evolution of the inference path, where our approach consistently recognized the value of initial steps, incorporating every layer from the first model (our PS-merged model), except for the last decoding layer and the embedding layer.
**Towards Reference-Free Metrics** Neural-based, reference-dependent metrics like COMET-22 demonstrate greater consistency with reference-free metrics and robustness compared to BLEU. For instance, with COMET-22, our models show significant improvements like other reference-free models over ALMA-13B-LoRA and comparable performance to GPT-4, e.g., 87.74 (Ours) vs. 87.68 (GPT-4) when en\(\rightarrow\)xx. However, it is important to note that, according to reference-free metrics, gold references are often inferior to system-generated translations, potentially indicating quality issues in the references that could impact COMET-22 evaluations. Consequently, inconsistencies still exist between COMET-22 and reference-free models like XCOMET. For example, XCOMET rates ALMA-R model on average higher than WMT winners (89.11 vs. 87.13), while COMET-22 favors WMT winners (85.21 vs. 85.60). In line with the recommendations in Freitag et al. (2023), we advocate for the use of reference-free models to circumvent the potential quality issues of references.
Both the third and fourth examples require the model to have good tabular understanding and reasoning ability. In the third example (about _Massif_), ChatQA-70B gives correct answer by comparing the size of preserved area against 3100 ha, while GPT-4 fails to do so. In the fourth example (about _John B. England_), ChatQA-70B correctly lists three dates but misses one date, while GPT-4 correctly answers the question.
where \(D\) is the preference dataset, \(x\) is the question, \(y_{w}\) and \(y_{l}\) are the chosen and rejected answers respectively, \(\theta\) is the learnable parameters of the model, and \(\gamma_{\pi}(x,y_{w},y_{l})=\log\frac{\pi(y_{w}|x)}{\pi(y_{l}|x)}\), _i.e.,_ the logratio of the chosen and rejected samples w.r.t. the policy \(\pi\).
**Conversational Performance.** Patient actors and specialist raters both evaluated AMIE's performance to be higher than PCPs on metrics related to empathy and communication skills. These axes comprised a majority of the dimensions that were evaluated. This general finding is consistent with a prior study where LLM responses were found to be more empathetic than the responses from clinicians to health questions posted on Reddit [73]. However, the findings in that study may not be generalised directly to our setting due to the differences in study design. Specifically, prior work has not involved a direct, randomised comparison of physicians and AI systems in a prospective simulation of multi-turn dialogue with the same patient. In both settings, the lack of voice-based and non-verbal visual communication may be an unfair disadvantage to clinicians.
\(\bullet\)_Environmental Feedback_. This feedback is obtained from the objective world or virtual environment. For instance, it could be the game's task completion signals or the observations made after the agent takes an action. In specific, ReAct [59] proposes constructing prompts using thought-act-observation triplets. The thought component aims to facilitate high-level reasoning and planning for guiding agent behaviors. The act represents a specific action taken by the agent. The observation corresponds to the outcome of the action, acquired through external feedback, such as search engine results. The next thought is influenced by the previous observations, which makes the generated plans more adaptive to the environment. Voyager [38] makes plans by incorporating three types of environment feedback including the intermediate progress of program execution, the execution error and self-verification results. These signals can help the agent to make better plans for the next action. Similar to Voyager, Ghost [16] also incorporates feedback into the reasoning and action taking processes. This feedback encompasses the environment states as well as the success and failure information for each executed action. SayPlan [31] leverages environmental feedback derived from a scene graph simulator to validate and refine its strategic formulations. This simulator is adept at discerning the outcomes and state transitions subsequent to agent actions, facilitating SayPlan's iterative recalibration of its strategies until a viable plan is ascertained. In DEPS [33], the authors argue that solely providing information about the completion of a task is often inadequate for correcting planning errors. Therefore, they propose informing the agent about the detail reasons for task failure, allowing them to more effectively revise their plans. LLM-Planner [60] introduces a grounded re-planning algorithm that dynamically updates plans generated by LLMs when encountering object mismatches and unattainable plans during task completion. Inner Monologue [61] provides three types of feedback to the agent after it takes actions: (1) whether the task is successfully completed, (2) passive scene descriptions, and (3) active scene descriptions. The former two are generated from the environments, which makes the agent actions more reasonable.
Retrievers play a crucial role in the RALM architecture. The information obtained through retrievers can significantly improve the accuracy of the LM. This section provides a summary of the retrieval methods commonly used in the RALM architecture. The retrieval methods are classified into four categories based on their methods and sources: Sparse Retrieval, Dense Retrieval, Internet Retrieval, and Hybrid Retrieval. Table 1 lists information about specific applications of retrievers in the RALM.
Monolingual TokenizerTable 3 demonstrates that the BPE-SP-33 tokenizer, on average, is the best-performing tokenizer, followed by the GPT-2 tokenizer. Interestingly, SentencePiece's implementation of BPE with a vocabulary size of 33k has been used for LLAMA2 (Touvron et al., 2023). Aggregated metrics provide a reasonable overview of the overall performance. However, it does not express potentially large performance differences across tasks. Therefore, we listed in Table 4 the obtained results for a list of selected tasks obtained by the best and worst performing tokenizer on this task. The results illustrate that the performance difference can be huge. For instance, for ARC-Easy, a commonsense reasoning task, the gap between the best and worst tokenizer is 9%.
Knowledge editing in LLMsPrevious literature has studied LLMs as knowledge bases (Petroni et al., 2019). In SS2.1, we aim to extend LLM knowledge bases through finetuning, as in Zhu et al. (2020).
Footnote 1: T5 was initially pretrained on the C4 dataset (about 750 GB), which is a cleansed dump of Common Crawl extracted from the web in April 2019.
In this section, we answer RQ1 of _how often an LLM performs the first hop of reasoning while pro cessing two-hop prompts_. We first introduce EntRec as a metric to approximate the LLM's internal recall of the bridge entity upon its descriptive mention in a prompt (SS5.1). Next, we propose to measure how often this recall increases when changing the input prompt to indirectly mention the bridge entity (SS5.2). Then, we evaluate this using TwoHopFact and answer RQ1 (SS5.3).
**ERNIE 3.0:** In [81], Sun et al. proposed a unified framework named ERNIE 3.0 for pre-training large-scale knowledge enhanced models. It fuses auto-regressive network and auto-encoding network, so that the trained model can be easily tailored for both natural language understanding and generation tasks using zero-shot learning, few-shot learning or fine-tuning. They have trained ERNIE 3.0 with 10 billion parameters on a 4TB corpus consisting of plain texts and a large-scale knowledge graph. Fig 17 illustrates the model architecture of Ernie 3.0.
Figure 2: **Overview of our RAFT method. The top-left figure depicts our approach of adapting LLMs to _reading_ solution from a set of positive and negative documents in contrast to standard RAG setup where models are trained based on the retriever outputs, which is a mixture of both memorization and reading.
Open Book ExamIn contrast, we liken the open-book exam setting to the scenario in which the LLM can refer to external sources of information (e.g., a website or a book chapter). In such scenarios, typically, the LLM is paired with a retriever which retrieves 'k' documents (or specific segments of the document) which are appended to the prompt. It is only through these documents retrieved that the LLM gains access to "new knowledge". As a result, we argue that the LLM's performance in these settings, where it is trained as a general-purpose LLM is largely dependent on the quality of the retriever and how accurately the retriever can identify the most relevant piece of information.
Property 2.1: _Gains increase with the number of steps._ We analyze the number of steps by isolating \(S\). We tune \(S\) from \(1\) to \(8\), while keeping the values of \(I\) and \(K\) constant across four groups of different values, ranging from small to large, respectively. Figure 6 (middle) shows that as the number of steps increases, there is a corresponding increase in performance gain. Additionally, we find that when \(I\) and \(K\) are increased (which indicates a higher difficulty), the performance gains are more significant, e.g., \(4\%\)-\(18\%\) gains over \(\{I=10,K=2\}\) compared to \(16\%\)-\(48\%\) over \(\{I=100,K=4\}\).
Other works focused on advancing agent patterns highlight that while breaking a larger problem into smaller subproblems can be effective at solving complex tasks, single agent patterns often struggle to complete the long sequence required [22; 6].
2. Preprocessing: Preprocess the data by removing noise, correcting grammar and spelling errors, and converting the text to a standard format. This step is crucial for the model to understand the context of the data.
**LLM Backbones.** We adopt LLaMA-based Touvron et al. (2023) and PaLM-based Anil et al. (2023) LLMs as our target LLMs in our experiments. For LLaMA-based target LLMs, we use Gemini-Pro Team et al. (2023) as the strong LLM, and LLaMA-7B, -13B as the target LLMs. For PaLM-based target LLMs, we use text-unicorn as the strong LLM, and text-boson as the target LLM. PaLM-based models and Gemini-Pro are accessible through Google Cloud API1.
**SFT Stage.** We find it still necessary to conduct a separate SFT phase. We utilize SFT data similar to the annealing phase excluding pre-training data and train with approximately 6 billion tokens. The learning rate for SFT is aligned with the one at the end of annealing, and a WSD Scheduler with exponential decay is also employed.
Evaluation results for applying sDPO to the SFT base model, along with results for other top-performing models are shown in Tab. 2. Comparing the pretrained-only 'SOLAR 10.7B' to the instruction-tuned 'SOLAR 10.7B + SFT', we can see an increase of \(+5.24\) in terms of H4. Applying sDPO on SOLAR 10.7B + SFT further increases the H4 score upto \(74.31\), an improvement of \(+4.80\). Notably, 'SOLAR 10.7B + SFT + sDPO' outperforms other larger models such as Mikral 8x7B-Instruct-v0.1, despite the smaller number of parameters. This highlights that effective alignment tuning could be the key to unlocking next level performance for smaller LLMs. Further, applying sDPO results in substantially higher score of \(72.45\) for TruthfulQA, which shows the effectiveness of the alignment tuning process.
This means that, to extract \(\mathbf{W}\) up to an orthogonal matrix, it's enough to recover some \(\mathbf{M}\) satisfying \(\|\mathbf{M}x_{i}\|=1\) for all \(i\). To compute \(\mathbf{M}\), we can actually recover the positive semi-definite \(\mathbf{A}\) satisfying Lemma G.1, which would give us a feasible \(\mathbf{M}\) by SVD or Cholesky decomposition.
The RAG preference rate is compared against two measurements: the prior probability and the deviation from the prior value. The former is computed by accessing the log probabilities from the OpenAI API call. As these are provided in log scale, we exponentiate them to produce linear probabilities when presenting the results. The latter is computed in several ways. For the Drug Dosages, Sports Statistics, and Latest News datasets, the absolute log fold change between the prior value and the modified value is computed; for the Wikipedia Dates dataset, the simple absolute year change is used; and for the Wikipedia Names and Locations datasets, each categorical change is presented in order of degree of modification.
Due to the huge number of model parameters, LLMs take a significant memory footprint for inference, making it very costly to be deployed in real-world applications. In this section, we discuss how to reduce the memory footprint of LLMs via a popular model compression approach (_i.e._, model quantization), so that large-sized LLMs can be used in resource-limited settings, which also likely reduces the inference latency.
While RAISE significantly improves upon existing methods in some respects, the researchers also highlighted several issues. First, RAISE struggles to understand complex logic, limiting its usefulness in many scenarios. Additionally, RAISE agents often hallucinated with respect to their roles or knowledge. For example, a sales agent without a clearly defined role might retain the ability to code in Python, which may enable them to start writing Python code instead of focusing on their sales tasks. These agents might also give the user misleading or incorrect information. This problem was addressed by fine-tuning the model, but the researchers still highlighted hallucination as a limitation in the RAISE implementation.
Footnote 44: Following [33], we mainly discuss symbolic reasoning tasks specially designed for evaluating LLMs. We do not consider symbolic reasoning methods in traditional NLP tasks, such as deducing logical rules from the knowledge graphs in KBQA.
In this work, we address this gap and conduct an extensive study in which we measure the impact of the tokenizer on the model performance.
Reverse training consists of taking a training dataset with \(N\) samples \(\{x_{1},\dots,x_{N}\}\) and constructing the set of reversed samples
FLARE (Jiang et al., 2023) uses StrategyQA, which contains a significant number of yes/no questions from a diverse range of sources. In addition, the authors request that the model provide the exact reasoning process and the final answer that determines the yes/no answer, ensuring that the answer matches the gold answer exactly. The incorporation of in-context samples into the retrieved content, along with training using data from the COPA, HellaSwag, and PIQA datasets, has resulted in the development of LLM-R (Wang et al., 2023) model that exhibits excellent performance. The fundamental concept of the ITER-RETGEN (Gao et al., 2022) model employs an iterative methodology to integrate retrievers and language models. The model was trained for the Commonsense Reasoning task using the StrategyQA dataset and achieved its optimal performance at seven iterations. In contrast, KG-BART (Shao et al., 2023) is designed to prioritize the Commonsense Reasoning task and employs knowledge graphs to enhance its performance in this area. This approach has proven effective in significantly improving the model's ability to complete the Commonsense Reasoning task, with performance approaching that of human beings under certain evaluation metrics.
For more details on RAG framework and its relevant works, we refer the readers to this survey of retrieval augmented generations [165].
Figure 6: Illustration of Result 8 (see Figure 14 for \(\mathsf{bioS}\) data and Figure 15 for \(\mathsf{bioD}\) data in Appendix C).
GPT2 was introduced in [26]. Due to its limitations from the absolute positional embedding [2], we adopt its modern variant, _rotary positional embedding_[7, 31], which we still refer to as GPT2 for convenience. Additionally, we disable dropout, which has been shown to improve performance in language models [33]. We explore a wide range of model sizes while using a fixed dimension-per-head of 64. The notation GPT2-\(\ell\)-\(h\) represents \(\ell\) layers, \(h\) heads, and \(64h\) dimensions; for example, GPT2-small corresponds to GPT2-12-12. The default GPT2Tokenizer is used, converting people's names and most attributes into tokens of variable lengths. In examining the impact of model architectures on scaling laws in Section 7, we will also use LLaMA/Mistral architectures [19, 32].
More details of training and computations.We use 4 Nvidia A100 with 80GB memory to train our models. All models are trained for 3 epochs with a batch size of 128, a peak learning rate of 2e-5 with 3% warmup steps, and linear decay afterward. We set the maximum token length to be 2,048 for the 7B model, and 1,524 for the 13B model due to the memory constraint. We use Deepspeed stage 3 (Rajbhandari et al., 2020) to conduct multi-GPU distributed training, with training precision Bfloat16 enabled. FlashAttention (Dao et al., 2022) is used to make the long-context training more efficient. We run inference of our trained models using 1-2 Quadro RTX 6000 GPUs with 24GB memory.
1. What if the response contains both unsafe and safe parts? Because the primary label is unsafe, if any part of the response is unsafe then mark the response as unsafe. 2. The response is not unambiguously safe, should it be labeled unsafe? No. Only mark it as unsafe if there are elements of the response that are actually unsafe. 3. The model response is garbled, hard to read or not an obvious reply, should I still annotate it? Yes! You should annotate everything for whether it is unsafe or not. Please flag all garbled and hard to read responses using the flags column.
Figure 8: The distribution of near-duplicate cluster sizes from running NearDup on each dataset.
We conduct ablation studies in terms of the number of unanswerable samples in HumanAnnotatedConvQA. We find that using a small amount of unanswerable samples (e.g., 1.5k) is able to achieve remarkable results on the unanswerable evaluation, as well as the overall conversational QA score (i.e., Avg-CQA). Interestingly, incorporating more unanswerable samples does not necessarily lead to higher accuracy scores in the unanswerable evaluation.
_Remark 1.6_.: Since 2bit/param is obtained after sufficient training, training longer _may not_ further improve model capacity, _but quantization can_. While not covered in this paper, our framework also provides a principled playground to compare different quantization methods. * Section 9: How sparsity (MoE) affects model capacity. Mixture-of-experts (MoE) models offer faster inference than dense models but often underperform dense models with the same total parameter count (not effective parameters). We show that this performance drop is likely not due to a lack of knowledge storage capability. * Result 9: MoE models, even with 32 experts, only reduce 1.3x in capacity compared to the base scaling laws, despite using just 8.8% of the total parameters during inference. * Section 10: How junk knowledge affects model capacity. Not all pretrain data are equally useful. Much of the internet data lacks valuable knowledge for training language models [24], while knowledge-rich sources like Wikipedia represent only a small fraction of the training tokens. We explore the impact on model capacity by conducting a controlled experiment with both useful and "junk" data. * Result 10+11: Junk data significantly reduces model capacity. As an example, with a 1:7 ratio of "useful to junk" training tokens, capacity for useful knowledge _loses by a factor of 20_x, even when useful knowledge is exposed 100 times.8 Footnote 8: The loss factor improves to 3x/1.5x/1.3x with 300/600/1000 exposures of useful knowledge, compared to Result 4 which involves training without junk for only 100 exposures. Result 12]An _effective mitigation_ is to prepend a special token to all useful knowledge. This is akin to adding a domain name like wikipedia.org at the start of every Wikipedia paragraph; the model _autonomously_ identifies high-quality data without prior knowledge of valuable domains. In the example above, the loss factor improves from 20x to 2x.
This update rule (\(Linear+Delta\)) leaves the associative matrix unmodified if the KV binding already exists in the memory while still tracking the same normalization term as the former one (\(Linear\)) for numerical stability.
**Capacity Leap**. Although GPT-2 is intended to be an "unsupervised multitask learner", it overall has an inferior performance compared with supervised fine-tuning state-of-the-art methods. Because it has a relatively small model size, it has been widely fine-tuned in downstream tasks, especially the dialog tasks [124, 125].
\(\bullet\)_These five closed-source models achieve promising results as general-purpose task solvers, in which ChatGPT mostly performs the best_. ChatGPT, Claude, Claude 2, Davinci003 and Davinci002 perform well in most of tasks, including complex tasks (_e.g.,_ GSM8k), which have shown great potential to be general-purpose task solvers. Among them, ChatGPT exhibits a more superior model capacity on the evaluation tasks, winning the most across all tasks. In some evaluation tasks, the performance gap between ChatGPT and other closed-source models is very large, especially for complex tasks _e.g.,_ 78.47 (ChatGPT) _vs._ 49.96 (Davinci002) on GSM8k, and 79.88 (ChatGPT) _vs._ 51.22 (Claude) on HumanEval.
+ Footnote †: This bandana makes the perfect addition to every für babies birthday collection! With its sparkly crown pattern, your pup will be ready for every birthday celebration! _With snaps for security, this bandana is made with love, down to the very last stitch!
For RealNews and LM1B, derived from news sites, we observe that many near-duplicates occur because the same news article appears on multiple news sites with slightly different formatting. For example, in LM1B, there is one example that starts "_MINEOLA, N.Y. - New York officials say_ [...]" and another that starts "_( AP ) - New York officials say_ [...]". The two examples are otherwise identical.
* We introduce _WizardMath_ model, which enhances the mathematical reasoning abilities for open-source pretrained large language model Llama-2 [20]. * We propose a new method, _Reinforcement Learning from Evol-Instruct Feedback_ (_**RLEIF**_)_, alongside _Evol-Instruct_ and Reinforcement Learning, for improving LLM reasoning performance. * _WizardMath_ surpasses all other open-source LLMs by a substantial margin in terms of mathematical reasoning, including Llama-2 70B [20], Llama-1 65B [4], Falcon-40B [21], MPT-30B [8], Baichuan-13B Chat [9] and ChatGLM2 12B [45] on both GSM8k [42] and MATH [43]. * _WizardMath_ significantly outperforms various main closed-source LLMs, such as ChatGPT [5], GPT-3.5, Claude Instant [39], PaLM-2 [44], PaLM-1 [7] and Minerva[15] on GSM8k.
* _Ranking-based approach._ In early work [367], human labelers often evaluate model-generated outputs in a coarse-grained manner (_i.e._, only selecting the best) without taking into account more fine-grained alignment criteria. Nonetheless, different labelers may hold diverse opinions on the selection of the best candidate output, and this method disregards the unselected samples, which may lead to inaccurate or incomplete human feedback. To address this issue, subsequent studies [116] introduce the Elo rating system to derive the preference ranking by comparing candidate outputs. The ranking of outputs serves as the training signal that guides the model to prefer certain outputs over others, thus inducing outputs that are more reliable and safer. * _Question-based approach._ Further, human labelers can provide more detailed feedback by answering certain questions designed by researchers [81], covering the alignment criteria as well as additional constraints for LLMs. Specially, in WebGPT [81], to assist the model in filtering and utilizing relevant information from retrieved documents, human labelers are required to answer questions with multiple options about whether the retrieved documents are useful for answering the given input. * _Rule-based approach._ Many studies also develop rule-based methods to provide more detailed human feedback. As a typical case, Sparrow [116] not only selects the response that labelers consider the best but also uses a series of rules to test whether model-generated responses meet the alignment criteria of being helpful, correct, and harmless. In this way, two kinds of human feedback data can be obtained: (1) the response preference feedback is obtained by comparing the quality of model-generated output in pairs, and (2) the rule violation feedback is obtained by collecting the assessment from human labelers (_i.e._, a score indicating to what extent the generated output has violated the rules). Furthermore, GPT-4 [46] utilizes a set of zero-shot classifiers (based on GPT-4 itself) as rule-based reward models, which can automatically determine whether the model-generated outputs violate a set of human-written rules.
With this equation, step accuracy can be estimated from the final answer accuracy. We experimented on GSM8K.
As Sophie continued to investigate the sphere, she began to notice strange things happening around her. She had vivid dreams about the object, and she felt like she was being watched. She even started to develop new abilities, like telekines and telepathy.
**Finance** is an important field where LLMs have promising application prospects. LLMs have been employed on various finance related tasks, such as numerical claim detection [922], financial sentiment analysis [923], financial named entity recognition [924], and financial reasoning [925]. Despite the competitive zero-shot performance exhibited by general-purpose LLMs in the finance tasks, they still underperform domain-specific PLMs containing million-scale parameters [922]. To leverage the scaling effect of LLMs, researchers collect large-scale finance corpora for continually pre-training LLMs (_e.g.,_ BloombergGPT [360], XuanYuan 2.0 [926], and FinGPT [927]). BloombergGPT has demonstrated remarkable performance across a diverse range of financial tasks while maintaining competitive performance in general-purpose tasks [360]. Nevertheless, it is imperative to consider the potential risks in the application of LLMs in finance, as the generation of inaccurate or harmful content by LLMs could have significant adverse implications for financial markets [360]. Therefore, it needs more strict reviewing and monitoring on the use of LLMs in the financial field.
According to a 2023 article, Pearson Education, a UK-based educational publisher, reported the highest revenue in 2021, with Wiley and McGraw Hill being the top two US-based educational publishers in terms of revenue.33
Moreover, while we see a clear scaling trend with the first hop of the latent multi-hop reasoning pathway with increasing model size, we do not see such scaling evidence for the second-hop reasoning pathway. This could be the reason behind the observation of Ofir Press et al. (2023) that the compositionality gap (the ratio of how often models can correctly answer all sub-problems but not generate the overall solution) does not decrease with increasing model size.
Gao et al., 2023) completely fails for our significantly more challenging setting. Given this evidence, we call for broadening the scope of error types and task settings (including our current task of book-length summarization) considered by current faithfulness evaluation benchmarks.
However, building a conversational QA model, that can match the accuracy of the state-of-the-art black-box model, i.e., GPT-4 (OpenAI, 2023), is still a grand challenge for the research community. In this work, we introduce ChatQA-70B, a white-box conversational QA model with GPT-4 level accuracy, through a proposed two-stage instruction tuning recipe, an enhanced retriever for retrieval-augmented generation (RAG) in conversational QA, and careful data curation process.
In our DFS merging experiments, \(M=64,r=3\), and consequently, \(T=M\times r=192\). We kept the last 200 examples in the training data as our validation set and optimize on the rest of the data with a batch size of 200. We report the performance of the snapshot that achieved the highest accuracy in the validation set, and the test set is strictly isolated from the optimization process. We adopted CMA-ES in EvoJAX [46], it optimized \(\mathcal{I}\) and \(W\) for a total of 100 generations with a popluation size of 128, and we used the default hyper-parameters. We limited our DFS merging to two models \(A\) and \(B\) to ensure that the final model remains modest in size and can be run on a single GPU, but in principle, the methodology can scale to merging multiple models. During the merging, model \(A\)'stokenizer, and input/output embeddings are utilized. Furthermore, to maintain compatibility with the embedding layers, we mandate that the initial and final transformer layers of model \(A\) define the start and the end of the inference path. We initialized the indicator array \(\mathcal{I}\) so that all layers in model \(A\) are more likely to be included as initial hops in the inference path to shorten the search time.
**Stop Using BLEU** BLEU, a metric extensively utilized for decades, often diverges from neural-based and reference-free metrics, a phenomenon also observed in previous studies (Xu et al., 2023; Freitag et al., 2023). For instance, WMT competition winners often exhibit superior performance according to BLEU (or COMET-22), yet this is not corroborated by reference-free models. A case in point is the WMT winners scoring an exceptionally high 64.14 BLEU in cs\(\rightarrow\)en translations, significantly outperforming other models by 20 BLEU points. However, reference-free evaluations suggest these translations are inferior to those generated by our models and GPT-4. We hypothesize that this discrepancy may arise from WMT models being trained on domain-specific data closely related to the WMT test set, leading to high lexical matches but lacking semantic depth as evaluated by neural-based metrics. While BLEU scores are effective for assessing basic functionality in weaker models, their utility diminishes with advanced translation models capable of generating diverse translations. In such contexts, relying solely on BLEU for evaluation appears increasingly outdated.
Figure 1: **Inconsistent knowledge in GPT-4.** GPT-4 correctly gives the name of Tom Cruise’s mother (left). Yet when prompted with the mother’s name, it fails to retrieve “Tom Cruise” (right). We hypothesize this ordering effect is due to the Reversal Curse. Models trained on “A is \(B\)” (e.g. “Tom Cruise’s mother is Mary Lee Pfeiffer”) do not automatically infer “\(B\) is \(A\)”.
We evaluate the performance of intermediate training checkpoints. The checkpoints are saved around every 43B tokens during pre-training. We plot the points of task performance (\(y\)-axis) and training loss (\(x\)-axis) in Figure 1. From the curves, we can see that the training loss is a good predictor of the performance on 12 downstream tasks.
* **An "inner" self-play loop** where AMIE leveraged in-context critic feedback to refine its behavior on simulated conversations with an AI patient agent. * **An "outer" self-play loop** where the set of refined simulated dialogues were incorporated into subsequent fine-tuning iterations. The resulting new version of AMIE could then participate in the inner loop again, creating a continuous learning cycle.
In this work, we introduce Self-Instruct, a semi-automated process for instruction-tuning a pretrained LM using instructional signals from the model itself. The overall process is an iterative bootstrapping algorithm (see Figure 2), which starts off with a limited (e.g., 175 in our study) seed set of manually-written tasks that are used to guide the overall generation. In the first phase, the model is prompted to generate instructions for new tasks. This step leverages the existing collection of instructions to create more broad-coverage instructions that define (often new) tasks. Given the newly-generated set of instructions, the framework also creates input-output instances for them, which can be later used for supervising the instruction tuning. Finally, various heuristics are used to automatically filter low-quality or repeated instructions, before adding the remaining valid tasks to the task pool. This process can be repeated for many iterations until reaching a large number of tasks.
**MATH**Hendrycks et al. (2021). This dataset focuses on competitive-level math problems that requires high levels of reasoning ability and mathematical knowledge. It consists of 7,500 training examples and 5,000 test cases. We use the training examples to generate synthetic data.
In order to renew the internal knowledge of LMs, one might consider pretraining LMs from scratch with a newly updated text corpus of a scale similar to the one used during initial pretraining, such as a recent dump of the entire Wikipedia. However, this approach is computationally demanding and also environmentally harmful (Patterson et al., 2021). Another alternative approach is continuing the pretraining process on a much smaller corpus containing new world knowledge, but such a methodology is known to suffer from _catastrophic forgetting_(McCloskey and Cohen, 1989; Kirkpatrick et al., 2017), where the models forget previously learned knowledge as they acquire new knowledge.
Motivated by the challenges of applying reinforcement learning algorithms on large-scale problems such as fine-tuning language models, our goal is to derive a simple approach for policy optimization using preferences directly. Unlike prior RLHF methods, which learn a reward and then optimize it via RL, our approach leverages a particular choice of reward model parameterization that enables extraction of its optimal policy in closed form, without an RL training loop. As we will describe next in detail, our key insight is to leverage an analytical mapping from reward functions to optimal policies, which enables us to transform a loss function over reward functions into a loss function over policies. This change-of-variables approach avoids fitting an explicit, standalone reward model, while still optimizing under existing models of human preferences, such as the Bradley-Terry model. In essence, the policy network represents both the language model and the (implicit) reward.
Figure 14: The prompt to judge the correctness of middle steps for GPT-4. We utilize the version of gpt-4-1106-preview. To help the prediction of GPT-4, we provide the groundtruth in the prompt for reference.
Figure 1: Illustration of continued pre-training (first row), continued pre-training followed by instruction-tuning (second row), and pre-instruction-tuning before continued pre-training (last row), along with their accuracies on evaluation questions. Each right-pointing light-blue triangle indicates a training phase.
U-PaLM is later instruction-finetuned as Flan-PaLM [74]. Compared to other instruction finetuning work mentioned above, Flan-PaLM's finetuning is performed using a much larger number of tasks, larger model sizes, and chain-of-thought data. As a result, Flan-PaLM substantially outperforms previous instruction-following models. For instance, Flan-PaLM-540B, which is instruction-finetuned on 1.8K tasks, outperforms PaLM-540B by a large margin (+9.4% on average). The finetuning data comprises 473 datasets, 146 task categories, and 1,836 total tasks, as illustrated in Fig 14.
In order to accurately emulate the ratings of specialists on the clinical criteria in our OSCE evaluation framework, we developed a model-based auto-evaluation procedure leveraging the AMIE model to score dialogues from 1 to 5 based on how well they exemplified those qualitative criteria. We initially focused on a subset of four clinical axes from the PACES criteria (see Table A.2), however, this procedure can be easily extended to other ratings.
Our URL filtering focuses on finding domains that are related to adult content, that may be harmful to users, or that are very likely to contain mostly unstructured text/spam (e.g., file hosting websites). First, we aggregated a list of 4.6M domains, detailed in Appendix G.1.1, that we explicitly ban; then, we built a simple URL scoring system, based on matching subwords in the URL against a list of words we curated (see Appendix G.1.2). We curated this list of words based on manual inspection, cross-referencing results with pages surfaced by ToxicBERT as being outliers in toxicity (Hanu and Unitary team, 2020).
* We discover the quantitative predictability of model performance regarding data mixture, and summarize this into a functional relationship, namely the data mixing laws. * We propose a pipeline to predict model performance of large-scale training on different mixture proportions but only experiments on small models with few training data through nested use of scaling laws of training steps, model sizes, and data mixing laws. * We experiment to verify the reliability of our data mixing laws and prediction pipeline, showing its effectiveness in optimizing model performance, balancing model capabilities, and the prospects of guiding the design of the data schedule.
**Efficient attention.** The efficient attention techniques attempt to improve the efficiency of the dot-product attention with an approximation or a system-level optimization. Multiple directions have been explored for different forms of efficient attention approximation, including sparsity-based (Child et al., 2019; Beltagy et al., 2020; Sukhbaatar et al., 2021; Ding et al., 2023) and linear attention approximation (Shen et al., 2018; Katharopoulos et al., 2020; Schlag et al., 2021). Among those, the linear attention variants are closely related to the associative memory matrix (Schlag et al., 2020; Schlag et al., 2021) and the metelearned neural memory (Munkhdalai et al., 2019), where KV bindings (Smolensky, 1990) are stored in Fast-Weights (Hinton and Plaut, 1987; Schmidhuber, 1992; Ba et al., 2016) that are modified in with respect to new contextual information. More recently, system-level optimization techniques have been proposed by leveraging specific hardware architecture to make the exact attention computation more efficient (Dao et al., 2022; Liu et al., 2023).
Measuring Update of Outdated World KnowledgeIn this work, we define _outdated_ world knowledge as information that is conflicting between \(D_{0}\) and \(D_{1}\). For example, the President of the US may be _Barack Obama_ in \(D_{0}\) and _Joe Biden_ in \(D_{1}\). In this case, the LM should update its internal knowledge as _Joe Biden_ as the US president. If an LM is pretrained on both \(D_{0}\) and \(D_{1}\) simultaneously, there is no guarantee that the LM will acquire the correct, recent information from \(D_{1}\), especially in cases where \(D_{0}\) is much bigger than \(D_{1}\), which is one of the biggest difference between the CKL and traditional CL setting. For measuring _update_ of outdated information, we construct UpdatedLAMA which is made up of cloze statements for which answers can be found in both \(D_{0}\) and \(D_{1}\), but are conflicting.
We can draw an analogy to an open-book exam. Existing in-context retrieval methods are equivalent to taking an open-book exam without studying. Alternatively, existing fine-tuning based approaches implement "studying" by either directly "memorizing" (Xiong et al., 2023) the input documents or answering practice questions (Wang et al., 2022) without referencing the documents. While these approaches leverage in-domain learning they fail to prepare for the open-book nature of test setting.
We fine-tuned three versions of FlanT5 (small, base, and large) (Chung et al., 2022) model. Our fine-tuning experiments encompassed both full parameter tuning and PEFT. Our focus was particularly on the QLoRA method (Dettmers et al., 2023), which is widely recognized for its approach of keeping the pre-trained model parameters fixed while incorporating trainable rank decomposition matrices into each layer of the Transformer architecture. Despite experimenting with various numbers of training epochs, we observed that the results plateaued after a certain point. Table 5 included presents our finalized hyperparameters for fine-tuning.
If the Jaccard Index between \(d_{i}\) and \(d_{j}\) is sufficiently high, it is likely that documents are approximate matches of each other. To efficiently approximate the Jaccard index, MinHash constructs document signatures by sorting each of the \(n\)-grams via a hash function, and then keeping only the \(k\) smallest hashed \(n\)-grams. There are multiple ways to construct estimators of the Jaccard index from these kinds of signatures [24].
Total number of gems = Number of diamonds + Number of rubies + Number of emeralds = 175 + 140 + 280 = 605.
RegularizationElastic Weight Consolidation (EWC) (Kirkpatrick et al., 2017) is a method that regularizes important parameters of previous tasks while training for the current tasks, helping mitigate _the negative backward transfer_ of previous tasks. Important parameters are measured via a Fisher information matrix computed by measuring the magnitude of the gradient update step of each parameter during training of previous tasks.
The specific de-duplication method can vary based on the nature of the data and the requirements of the particular language model being trained. It may involve comparing entire data points or specific features to identify and eliminate duplicates. At the document level, existing works mainly rely on the overlap ratio of high-level features (e.g. n-grams overlap) between documents to detect duplicate samples.
**Experiment.** We experiment with approach (ii) and (iii) above. For (ii), we prompt GPT-3.5-Turbo and Mixtral-8x7B-DPO for relevance rating with the same prompt template (Figure 14). For (iii), we use OpenAI's text-embedding-3-large as the embedding model.
Figure 2: **Overview of randomized study design. A primary care physician (PCP) and AMIE perform (in a randomized order) a virtual remote Objective Structured Clinical Examination (OSCE) with simulated patients via online multi-turn synchronous text chat and produce answers to a post-questionnaire.
To assess the quality of the model's output, we use the following prompt to GPT-4 as a benchmark. We note that we adapt this prompt to the target language by translating it and adjusting the {target language} accordingly.
Researchers typically consider the timing of retrieval in two situations: when working on tasks that require multiple retrievals, such as long dialogue generation and multi-hop problems, or when it is impossible to find a suitable and relevant document. Using irrelevant documents can harm the accuracy of the output.
**Other Practical Tricks.** In practice, there are also several useful strategies and tricks that are helpful to improve the fine-tuning performance of LLMs. We list several representative ones as follows:
* As shown in Fig. 5(a) and Fig. 5(b), among all cases where LLMs have minimized perplexity on documents, cases trained with more epochs or larger learning rates typically exhibit superior QA performance. We hypothesize that _more aggressive training leads to less overfitting to deceptive patterns in documents and better generalization when responding to questions_.