OpenAI

Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, Dario Amodei

Recent years have featured a trend towards pre-trained language representations in NLP systems, applied in increasingly ?flexible and task-agnostic ways for downstream transfer. First, single-layer representations were learned using word ?vectors [MCCD13, PSM14] and fed to task-specific architectures, then RNNs with multiple layers of representations ?and contextual state were used to form stronger representations [DL15, MBXS17, PNZtY18] (though still applied to ?task-specific architectures), and more recently pre-trained recurrent or transformer language models [VSP+17] have ?been directly fine-tuned, entirely removing the need for task-specific architectures [RNSS18, DCLT18, HR18].

This last paradigm has led to substantial progress on many challenging NLP tasks such as reading comprehension, ?question answering, textual entailment, and many others, and has continued to advance based on new architectures ?and algorithms [RSR+19, LOG+19, YDY+19, LCG+19]. However, a major limitation to this approach is that while ?the architecture is task-agnostic, there is still a need for task-specific datasets and task-specific fine-tuning: to achieve ?strong performance on a desired task typically requires fine-tuning on a dataset of thousands to hundreds of thousands ?of examples specific to that task. Removing this limitation would be desirable, for several reasons. ?
First, from a practical perspective, the need for a large dataset of labeled examples for every new task limits the ?applicability of language models. There exists a very wide range of possible useful language tasks, encompassing ?anything from correcting grammar, to generating examples of an abstract concept, to critiquing a short story. For many ?of these tasks it is difficult to collect a large supervised training dataset, especially when the process must be repeated ?for every new task.

對解決這些問題的一個潛在的路線是meta-learning1——在語言的上下文模型意味著模型發展廣泛技能的訓練時間和模式識別能力,然后使用這些能力在推理時迅速適應或識別所需的任務(見圖1.1)。最近的工作[RWC + 19]試圖做到這一點通過我們稱之為“語境學習”,使用文本輸入pretrained語言模型作為一種任務規范:模型條件在自然語言指令和/或一些示威活動的任務,然后將完成進一步的實例任務只需預測接下來會發生什么。雖然它顯示出了一些最初的希望，但這種方法取得的效果仍遠不及微調——例如[RWC+19]在自然問題上僅取得4%的成績，甚至它的55 F1 CoQa結果現在也落后于最先進的水平35分以上。元學習顯然需要大量的改進，才能成為解決語言任務的可行的實用方法。語言建模的另一個最新趨勢可能提供了一個前進的方向。近年來，transformer語言模型的容量大幅增加，從1億個參數[RNSS18]，到3億個參數[DCLT18]，再到15億個參數[RWC+19]，再到80億個參數[SPP+19]， 110億個參數[RSR+19]，最后是170億個參數[Tur20]。每一次增加都帶來了文本合成和/或下游NLP任務的改進，有證據表明，與許多下游任務相關的日志丟失隨著規模的增大呈現平穩的改善趨勢[KMH+20]。由于內環境學習涉及在模型的參數內吸收許多技能和任務，因此內環境學習能力可能會隨著規模的增長而顯示出類似的強勁增長，這是合理的。

?

• Fine-Tuning (FT) has been the most common approach in recent years, and involves updating the weights of a pre-trained model by training on a supervised dataset specific to the desired task. Typically thousands to hundreds of thousands of labeled examples are used. The main advantage of fine-tuning is strong performance on many benchmarks. The main disadvantages are the need for a new large dataset for every task, the potential for poor generalization out-of-distribution [MPL19], and the potential to exploit spurious features of the training data [GSL+18, NK19], potentially resulting in an unfair comparison with human performance. In this work we do not fine-tune GPT-3 because our focus is on task-agnostic performance, but GPT-3 can be fine-tuned in principle and this is a promising direction for future work.
• Few-Shot (FS) is the term we will use in this work to refer to the setting where the model is given a few demonstrations of the task at inference time as conditioning [RWC+19], but no weight updates are allowed. As shown in Figure 2.1, for a typical dataset an example has a context and a desired completion (for example an English sentence and the French translation), and few-shot works by giving K examples of context and completion, and then one final example of context, with the model expected to provide the completion. We typically set K in the range of 10 to 100 as this is how many examples can fit in the model’s context window (nctx = 2048). The main advantages of few-shot are a major reduction in the need for task-specific data and reduced potential to learn an overly narrow distribution from a large but narrow fine-tuning dataset. The main disadvantage is that results from this method have so far been much worse than state-of-the-art fine-tuned models. Also, a small amount of task specific data is still required. As indicated by the name, few-shot learning as described here for language models is related to few-shot learning as used in other contexts in ML [HYC01, VBL+16] – both involve learning based on a broad distribution of tasks (in this case implicit in the pre-training data) and then rapidly adapting to a new task.
• One-Shot (1S) is the same as few-shot except that only one demonstration is allowed, in addition to a natural language description of the task, as shown in Figure 1. The reason to distinguish one-shot from few-shot and zero-shot (below) is that it most closely matches the way in which some tasks are communicated to humans. For example, when asking humans to generate a dataset on a human worker service (for example Mechanical Turk), it is common to give one demonstration of the task. By contrast it is sometimes difficult to communicate the content or format of a task if no examples are given.
• Zero-Shot (0S) is the same as one-shot except that no demonstrations are allowed, and the model is only given a natural language instruction describing the task. This method provides maximum convenience, potential for robustness, and avoidance of spurious correlations (unless they occur very broadly across the large corpus of pre-training data), but is also the most challenging setting. In some cases it may even be difficult for humans to understand the format of the task without prior examples, so this setting is in some cases “unfairly hard”. For example, if someone is asked to “make a table of world records for the 200m dash”, this request can be ambiguous, as it may not be clear exactly what format the table should have or what should be included (and even with careful clarification, understanding precisely what is desired can be difficult). Nevertheless, for at least some settings zero-shot is closest to how humans perform tasks – for example, in the translation example in Figure 2.1, a human would likely know what to do from just the text instruction.

• 微調(FT)是近年來最常見的方法，它包括通過對特定于預期任務的監督數據集進行訓練來更新預訓練模型的權重。通常使用成千上萬的帶標簽的例子。調優的主要優點是在許多基準測試中具有強大的性能。主要缺點是每個任務都需要一個新的大數據集，可能會出現泛化不均勻分布[MPL19]，可能會利用訓練數據的虛假特征[GSL+18, NK19]，可能會導致與人類性能的不公平比較。在這項工作中，我們沒有對GPT-3進行微調，因為我們關注的是任務不確定性能，但是GPT-3原則上可以進行微調，這是未來工作的一個有希望的方向。
• 少量射擊(FS)是我們將在這項工作中使用的術語，用來指在推理時給模型一些任務的演示作為條件[RWC+19]，但不允許權重更新。如圖2.1所示,一個典型的數據集實例有一個上下文和所需的完成(例如一個英語句子翻譯和法國),和few-shot作品給K上下文和完成的例子,然后最后一個例子的情況下,模型將提供完成。我們通常將K設置在10到100之間，因為這是模型上下文窗口所能容納的示例數(nctx = 2048)。few-shot的主要優點是大大減少了對特定任務數據的需求，并減少了從一個大而窄的微調數據集學習過窄分布的可能性。這種方法的主要缺點是，到目前為止，其結果遠不如最先進的微調模型。此外，仍然需要少量特定于任務的數據。顯示的名字,few-shot學習作為語言模型與這里描述few-shot學習用于其他上下文毫升(HYC01輪式偵察車+ 16)-包括基于分布廣泛的學習任務(在本例中隱含在訓練的數據),然后迅速適應新任務。
• one -shot (1S)與few-shot相同，只是除了任務的自然語言描述之外，只允許進行一次演示，如圖1所示。區分“一步走”、“少一步走”和“零一步走”(見下圖)的原因是，“一步走”與某些任務傳達給人類的方式最接近。例如，當要求人類在人工工作服務(例如Mechanical Turk)上生成數據集時，通常會給出任務的演示。相比之下，如果沒有例子，有時很難傳達任務的內容或格式。
• Zero-Shot(0)與one-shot相同，只是不允許演示，并且只給模型一個描述任務的自然語言指令。這種方法提供了最大的便利性、潛在的魯棒性和避免虛假相關性(除非它們在大量的訓練前數據語料庫中廣泛出現)，但也是最具挑戰性的設置。在某些情況下，如果沒有之前的例子，人類甚至很難理解任務的格式，所以在某些情況下，這種設置是“不公平的困難”。例如,如果有人要求“讓200米短跑的世界紀錄表”,這個請求可以模糊,因為它可能不清楚什么格式表應該有或應該包括什么(甚至仔細澄清,需要理解恰恰是很困難的)。不過，至少在某些設置中，zero-shot最接近于人類執行任務的方式——例如，在圖2.1中的翻譯示例中，人類可能僅通過文本指令就知道該做什么。

Figure 2.1 shows the four methods using the example of translating English to French. In this paper we focus on zero-shot, one-shot and few-shot, with the aim of comparing them not as competing alternatives, but as different problem settings which offer a varying trade-off between performance on specific benchmarks and sample efficiency. We especially highlight the few-shot results as many of them are only slightly behind state-of-the-art fine-tuned models. Ultimately, however, one-shot, or even sometimes zero-shot, seem like the fairest comparisons to human performance, and are important targets for future work.

Sections 2.1-2.3 below give details on our models, training data, and training process respectively. Section 2.4 discusses the details of how we do few-shot, one-shot, and zero-shot evaluations.

我們使用與GPT-2 [RWC+19]相同的模型和架構，包括修改的初始化、預歸一化和其中描述的可逆標記，但我們在變壓器的層中使用交替密集和局部帶狀稀疏注意模式，類似于稀疏變壓器[CGRS19]。為了研究ML性能對模型大小的依賴關系，我們訓練了8種不同大小的模型，從1.25億個參數到1750億個參數的3個數量級，最后一個是我們稱為GPT-3的模型。先前的研究[KMH+20]表明，在有足夠的訓練數據的情況下，驗證損失的比例應近似于一個平滑的冪律，該冪律是大小的函數;許多不同大小的訓練模型允許我們測試驗證丟失和下游語言任務的假設。

?

Datasets for language models have rapidly expanded, culminating in the Common Crawl dataset2 [RSR+19] constituting nearly a trillion words. This size of dataset is sufficient to train our largest models without ever updating on the same sequence twice. However, we have found that unfiltered or lightly filtered versions of Common Crawl tend to have lower quality than more curated datasets. Therefore, we took 3 steps to improve the average quality of our datasets: (1) we downloaded and filtered a version of CommonCrawl based on similarity to a range of high-quality reference corpora, (2) we performed fuzzy deduplication at the document level, within and across datasets, to prevent redundancy and preserve the integrity of our held-out validation set as an accurate measure of overfitting, and (3) we also added known high-quality reference corpora to the training mix to augment CommonCrawl and increase its diversity.

?

Details of the first two points (processing of Common Crawl) are described in Appendix A. For the third, we added several curated high-quality datasets, including an expanded version of the WebText dataset [RWC+19], collected by scraping links over a longer period of time, and first described in [KMH+20], two internet-based books corpora (Books1 and Books2) and English-language Wikipedia.

用于語言模型的數據集已經迅速擴展，最終達到了常見的爬行數據集dataset2 [RSR+19]，總計近一萬億字。這樣大的數據集足以訓練我們最大的模型，而無需對同一序列進行兩次更新。然而，我們發現未過濾或輕度過濾版本的普通爬行往往比更有組織的數據集質量更低。因此，我們采取了3個步驟來提高數據集的平均質量:(1)我們下載和過濾的一個版本CommonCrawl基于相似性的一系列高品質參考全集,(2)我們在文檔級別執行模糊重復數據刪除,在和整個數據集,以防止冗余和保存我們伸出的完整性驗證設置為一個精確的衡量過度擬合,和(3)我們還添加了高質量的參考語料訓練增加CommonCrawl和增加其多樣性。

前兩個點的詳細信息(處理常見的爬行)描述在附錄a。第三,我們添加了幾個策劃高質量的數據集,包括WebText數據集的擴展版本(RWC + 19),收集的抓取鏈接在更長一段時間,和第一(公里/小時+ 20)中描述的兩個網絡書全集(Books1和Books2)和英文維基百科。

對于少彈學習，我們從任務的訓練集中隨機抽取K個樣本作為條件，根據任務的不同用1或2個新行分隔，以此來評估評估集中的每個樣本。對于LAMBADA和Storycloze，沒有可用的監督訓練集，所以我們從開發集中提取條件設置示例，并在測試集上進行評估。對于Winograd(原始的，不是超級膠水版本)，只有一個數據集，所以我們直接從它提取條件設置示例。

K can be any value from 0 to the maximum amount allowed by the model’s context window, which is nctx = 2048 for all models and typically fits 10 to 100 examples. Larger values of K are usually but not always better, so when a separate development and test set are available, we experiment with a few values of K on the development set and then run the best value on the test set. For some tasks (see Appendix G) we also use a natural language prompt in addition to (or for K = 0, instead of) demonstrations.

On tasks that involve choosing one correct completion from several options (multiple choice), we provide K examples of context plus correct completion, followed by one example of context only, and compare the LM likelihood of each completion. For most tasks we compare the per-token likelihood (to normalize for length), however on a small number of datasets (ARC, OpenBookQA, and RACE) we gain additional benefit as measured on the development set by normalizing by the unconditional probability of each completion, by computing P (completion|context) P (completion|answer context) , where answer context is the string "Answer: " or "A: " and is used to prompt that the completion should be an answer but is otherwise generic.

K可以是0到模型上下文窗口允許的最大數量之間的任何值，即nctx = 2048，適用于所有模型，通常適合10到100個示例。更大的K值通常但不總是更好的,所以當一組獨立的開發和測試是可用的,我們嘗試幾值K的開發設置,然后運行測試集上的最佳值。對于某些任務(參見附錄G)我們也使用自然語言提示除了(或K = 0,而不是)示威活動。

對于涉及從多個選項(多項選擇)中選擇一個正確完成的任務，我們提供了K個上下文示例加上正確完成，然后只提供一個上下文示例，并比較每個完成的LM可能性。對于大多數任務我們比較每個令牌的可能性(規范化長度),然而在少量的數據集(弧、OpenBookQA和比賽)我們獲得更多利益衡量發展設定的正常化的無條件概率每完成,通過計算P(完成|上下文)(完成|回答上下文),在回答上下文字符串“回答:”或“:”和用于提示完成應該答案但否則通用。

On tasks that involve binary classification, we give the options more semantically meaningful names (e.g. “True” or “False” rather than 0 or 1) and then treat the task like multiple choice; we also sometimes frame the task similar to what is done by [RSR+19] (see Appendix G) for details.

On tasks with free-form completion, we use beam search with the same parameters as [RSR+19]: a beam width of 4 and a length penalty of α = 0.6. We score the model using F1 similarity score, BLEU, or exact match, depending on what is standard for the dataset at hand.

Final results are reported on the test set when publicly available, for each model size and learning setting (zero-, one-, and few-shot). When the test set is private, our model is often too large to fit on the test server, so we report results on the development set. We do submit to the test server on a small number of datasets (SuperGLUE, TriviaQA, PiQa) where we were able to make submission work, and we submit only the 200B few-shot results, and report development set results for everything else.

對于涉及二分類的任務，我們給選項以語義上更有意義的名稱(例如“真”或“假”，而不是0或1)，然后把任務當作多項選擇;我們有時也會類似于[RSR+19]所完成的任務(詳見附錄G)。

對于自由形式完成的任務，我們使用與[RSR+19]相同的參數進行波束搜索:波束寬度為4，長度罰值為radial = 0.6。我們使用F1相似度評分、BLEU或精確匹配來給模型評分，這取決于手頭數據集的標準。

對于每個模型的大小和學習設置(0 -，1 -，和小樣本)，最終的結果會在測試集上公布。當測試集是私有的,我們的模型往往是太大,以適應在測試服務器上,所以我們報告的結果發展。我們提交到測試服務器上少量的數據集(超強力膠水,TriviaQA PiQa)我們能夠提交工作,我們只有200 b few-shot提交結果,并報告發展為一切設置結果。

In Figure 3.1 we display training curves for the 8 models described in Section 2. For this graph we also include 6 ?additional extra-small models with as few as 100,000 parameters. As observed in [KMH+20], language modeling ?performance follows a power-law when making efficient use of training compute. After extending this trend by two ?more orders of magnitude, we observe only a slight (if any) departure from the power-law. One might worry that these ?improvements in cross-entropy loss come only from modeling spurious details of our training corpus. However, we will ?see in the following sections that improvements in cross-entropy loss lead to consistent performance gains across a ?broad spectrum of natural language tasks. ?

Below, we evaluate the 8 models described in Section 2 (the 175 billion parameter parameter GPT-3 and 7 smaller ?models) on a wide range of datasets. We group the datasets into 9 categories representing roughly similar tasks. ?

在圖3.1中，我們展示了第2節中描述的8個模型的訓練曲線。在這個圖中，我們還包括了6個額外的超小型模型，這些模型只有100,000個參數。正如在[KMH+20]中觀察到的，在高效使用訓練計算時，語言建模性能遵循冪律。在將這一趨勢擴展兩個數量級之后，我們只觀察到與冪律有輕微的背離。人們可能會擔心這些交叉熵損失的改進僅僅來自于我們訓練語料庫的虛假細節建模。然而，在接下來的章節中，我們將看到交叉熵損失的改進可以在廣泛的自然語言任務中帶來一致的性能提升。 

下面，我們在廣泛的數據集上評估第2節中描述的8個模型(1750億參數GPT-3和7個較小的模型)。我們將數據集分成9個類別，這些類別代表大致相似的任務。 

在本節中，我們將測試GPT-3在傳統的語言建模任務以及相關任務上的性能，這些任務包括預測感興趣的單個單詞、完成句子或段落，或在可能完成的一段文本之間進行選擇。

The LAMBADA dataset [PKL+16] tests the modeling of long-range dependencies in text – the model is asked to predict the last word of sentences which require reading a paragraph of context. It has recently been suggested that the continued scaling of language models is yielding diminishing returns on this difficult benchmark. [BHT+20] reflect on the small 1.5% improvement achieved by a doubling of model size between two recent state of the art results ([SPP+19]?and [Tur20]) and argue that “continuing to expand hardware and data sizes by orders of magnitude is not the path forward”. We find that path is still promising and in a zero-shot setting GPT-3 achieves 76% on LAMBADA, a gain of 8% over the previous state of the art.

LAMBADA is also a demonstration of the flexibility of few-shot learning as it provides a way to address a problem that ?classically occurs with this dataset. Although the completion in LAMBADA is always the last word in a sentence, a ?standard language model has no way of knowing this detail. It thus assigns probability not only to the correct ending but ?also to other valid continuations of the paragraph. This problem has been partially addressed in the past with stop-word ?filters [RWC+19] (which ban “continuation” words). The few-shot setting instead allows us to “frame” the task as a ?cloze-test and allows the language model to infer from examples that a completion of exactly one word is desired. We ?use the following fill-in-the-blank format:

LAMBADA數據集[PKL+16]測試文本中遠程依賴的建?！Ｐ捅灰箢A測需要閱讀一段上下文的句子的最后一個單詞。最近有研究表明，語言模型的不斷擴大在這個困難的基準上產生的收益正在減少。[BHT+20]反思了在兩個最新的研究結果([SPP+19]和[Tur20])之間，模型尺寸增加了一倍，僅提高了1.5%，并認為“繼續以數量級擴展硬件和數據尺寸并不是前進的道路”。我們發現這條道路仍然很有希望，在零桿的情況下，LAMBADA的GPT-3實現了76%，比之前的技術水平提高了8%。
LAMBADA還演示了小樣本學習的靈活性，因為它提供了一種方法來解決這個數據集通常出現的問題。盡管LAMBADA中的完成總是一個句子的最后一個單詞，但是標準語言模型無法知道這個細節。因此，它不僅將概率分配給正確的結尾，也分配給其他有效的段落延續。這個問題已經部分解決了在過去的停止字過濾器[RWC+19](禁止“延續”字)。相反，few-shot設置允許我們將任務“設置”為一個cloze測試，并允許語言模型從示例中推斷出需要完成的恰好是一個單詞。我們使用以下填空格式:

?

When presented with examples formatted this way, GPT-3 achieves 86.4% accuracy in the few-shot setting, an increase ?of over 18% from the previous state-of-the-art. We observe that few-shot performance improves strongly with model ?size. While this setting decreases the performance of the smallest model by almost 20%, for GPT-3 it improves accuracy ?by 10%. Finally, the fill-in-blank method is not effective one-shot, where it always performs worse than the zero-shot ?setting. Perhaps this is because all models still require several examples to recognize the pattern.

One note of caution is that an analysis of test set contamination identified that a significant minority of the LAMBADA dataset appears to be present in our training data – however analysis performed in Section 4 suggests negligible impact on performance.

在本節中，我們將測量GPT-3回答有關廣泛事實知識的問題的能力。由于可能的查詢量巨大，這個任務通常是通過使用信息檢索系統查找相關文本，并結合學習生成給定問題和檢索文本的答案的模型來完成的。由于該設置允許系統搜索并對可能包含答案的文本進行條件設置，因此稱為“open-book”。[RRS20]最近證明，一個大型語言模型可以在不依賴輔助信息的情況下直接回答問題，表現得令人驚訝地好。他們將這種更嚴格的評估設置稱為“閉卷”。他們的工作表明，更高容量的模型可以表現得更好，我們用GPT-3測試了這一假設。我們在[RRS20]中的3個數據集上評估GPT-3: Natural Questions [KPR+19]、WebQuestions [BCFL13]和TriviaQA [JCWZ17]，使用相同的分割。注意，除了所有的結果都在閉卷設置中之外，我們使用的少樣本、一次小樣本和零小樣本的評估代表了比以前的閉卷QA工作更嚴格的設置:除了不允許外部內容外，也不允許對Q&A數據集本身進行微調。

GPT-3結果如表3.3所示。在TriviaQA上，我們在小樣本設置中達到了64.3%，在一小樣本設置中達到了68.0%，在小樣本設置中達到了71.2%。zero-shot result的表現已經比經過微調的T5-11B高出14.2%，而且在培訓前的問答時間跨度預測也比T5-11B高出3.8%。一次測試的結果提高了3.7%，與開放域QA系統的SOTA相匹配，該系統不僅進行了優化，而且利用了一種學習過的檢索機制，對包含21M文檔的15.3個參數密集向量索引進行檢索[LPP+20]。此外，GPT-3的少拍效果進一步提高了性能3.2%。
在網絡問題(WebQs)中，GPT-3在零桿設置中達到14.4%，在單桿設置中達到25.3%，在少桿設置中達到41.5%。相比之下，使用q&a特定的培訓前程序的優化T5-11B和優化T5-11B+SSM的比例分別為37.4%和44.7%。GPT-3在小樣本設置接近最先進的表現，微調模型。值得注意的是，與TriviaQA相比，WebQS從零桿到少桿的增益要大得多(事實上，WebQS的零桿和單桿性能都很差)，這可能表明WebQS的問題和/或它們的回答風格在GPT-3中是不分布的。然而，GPT-3似乎能夠適應這種分布，在少炮點的環境中恢復了良好的性能。

On Natural Questions (NQs) GPT-3 achieves 14.6% in the zero-shot setting, 23.0% in the one-shot setting, and 29.9% in ?the few-shot setting, compared to 36.6% for fine-tuned T5 11B+SSM. Similar to WebQS, the large gain from zero-shot ?to few-shot may suggest a distribution shift, and may also explain the less competitive performance compared to ?TriviaQA and WebQS. In particular, the questions in NQs tend towards very fine-grained knowledge on Wikipedia ?specifically which could be testing the limits of GPT-3’s capacity and broad pretraining distribution. ?

Overall, on one of the three datasets GPT-3’s one-shot matches the open-domain fine-tuning SOTA. On the other two ?datasets it approaches the performance of the closed-book SOTA despite not using fine-tuning. On all 3 datasets, we ?find that performance scales very smoothly with model size (Figure 3.3 and Appendix H Figure H.7), possibly reflecting ?the idea that model capacity translates directly to more ‘knowledge’ absorbed in the parameters of the model.

在自然問題(NQs)中，GPT-3在零桿設置中達到了14.6%，在單桿設置中達到了23.0%，在少桿設置中達到了29.9%，而在經過微調的T5 11B+SSM中達到了36.6%。與WebQS類似，從零桿到少桿的巨大增益可能意味著分布的轉移，這也可能解釋了與TriviaQA和WebQS相比競爭力較差的原因。特別是，NQs的問題傾向于維基百科上非常精細的知識，可以測試GPT-3的能力極限和廣泛的培訓前分布。 

總的來說，在三個數據集中的一個上，GPT-3的一次性匹配了開放域微調SOTA。在另外兩個數據集上，盡管沒有使用微調，它的性能接近封閉的SOTA。在所有3個數據集上，我們發現性能與模型大小的關系非常順利(圖3.3和附錄H圖H.7)，可能反映了模型容量直接轉化為更多吸收在模型參數中的“知識”的想法。

Winograd Schemas Challenge [LDM12]是NLP中的一項經典任務，當一個代詞在語法上有歧義，但在語義上對人來說沒有歧義時，該任務涉及確定該代詞指的是哪個詞。最近，經過微調的語言模型在原始Winograd數據集上取得了接近人類的性能，但是更困難的版本，比如反向挖掘的Winogrande數據集[SBBC19]，仍然顯著落后于人類的性能。我們測試了GPT-3在Winograd和Winogrande上的性能，通常是在零桿、一桿和少桿設置下。 
在Winograd上，我們使用[RWC+19]中描述的相同的“部分求值”方法，在原始的273個Winograd模式集上測試GPT-3。請注意，此設置與SuperGLUE基準中的WSC任務略有不同，后者以二進制分類的形式表示，需要提取實體來轉換為本節中描述的形式。Winograd的GPT-3在零桿、一桿和少桿設置中取得了88.3%、89.7%和88.6%的成績，沒有顯示出明確的上下文學習，但在所有情況下都取得了較好的成績，僅比最先進的和估計的人類性能低幾個點。我們注意到，污染分析在訓練數據中發現了一些Winograd模式，但這似乎只對結果有很小的影響(見第4節)。

?

ARC [CCE+18] is a dataset of multiple-choice questions collected from 3rd to 9th grade science exams. On the ?“Challenge” version of the dataset which has been filtered to questions which simple statistical or information retrieval ?methods are unable to correctly answer, GPT-3 achieves 51.4% accuracy in the zero-shot setting, 53.2% in the one-shot ?setting, and 51.5% in the few-shot setting. This is approaching the performance of a fine-tuned RoBERTa baseline ?(55.9%) from UnifiedQA [KKS+20]. On the “Easy” version of the dataset (questions which either of the mentioned ?baseline approaches answered correctly), GPT-3 achieves 68.8%, 71.2%, and 70.1% which slightly exceeds a fine-tuned ?RoBERTa baseline from [KKS+20]. However, both of these results are still much worse than the overall SOTAs ?achieved by the UnifiedQA which exceeds GPT-3’s few-shot results by 27% on the challenge set and 22% on the easy ?set. ?

On OpenBookQA [MCKS18], GPT-3 improves significantly from zero to few shot settings but is still over 20 points ?short of the overall SOTA. GPT-3’s few-shot performance is similar to a fine-tuned BERT Large baseline on the ?leaderboard. ?
Overall, in-context learning with GPT-3 shows mixed results on commonsense reasoning tasks, with only small and ?inconsistent gains observed in the one and few-shot learning settings for both PIQA and ARC, but a significant ?improvement is observed on OpenBookQA. GPT-3 sets SOTA on the new PIQA dataset in all evaluation settings.

Next we evaluate GPT-3 on the task of reading comprehension. We use a suite of 5 datasets including abstractive, ?multiple choice, and span based answer formats in both dialog and single question settings. We observe a wide spread ?in GPT-3’s performance across these datasets suggestive of varying capability with different answer formats. In general ?we observe GPT-3 is on par with initial baselines and early results trained using contextual representations on each ?respective dataset. ?

接下來我們對GPT-3進行閱讀理解任務的評估。在對話框和單一問題設置中，我們使用了一套5個數據集，包括抽象的、多項選擇和基于跨度的回答格式。我們觀察到GPT-3在這些數據集上的性能差異很大，這表明不同的回答格式具有不同的能力。一般來說，我們觀察到GPT-3與初始基線和使用上下文表示對每個各自數據集進行訓練的早期結果相同。?

GPT-3在CoQA [RCM19]自由形式會話數據集上表現最好(在人類基線的3個點內)，在QuAC [CHI+18]數據集上表現最差(低于ELMo基線13 F1)，該數據集需要建模結構化對話行為和師生交互的回答范圍選擇。下降(DWD + 19]數據集測試離散推理和計算能力在閱讀理解中,GPT-3在few-shot環境優于原始論文的BERT基線調整但仍遠低于人類的性能和先進的方法增強神經網絡與符號系統(RLL + 19)。在陣容2.0 [RJL18]上，GPT-3展示了它的少桿學習能力，與零桿設置相比提高了近10桿(69.8桿)。這使得它稍微優于原始論文中最好的微調結果。在RACE [LXL+17](一個針對初中和高中英語考試的多項選擇數據集)上，GPT-3的表現相對較弱，僅與最早使用上下文表示的研究相比具有競爭力，仍落后于SOTA 45%。

In order to better aggregate results on NLP tasks and compare to popular models such as BERT and RoBERTa in a ?more systematic way, we also evaluate GPT-3 on a standardized collection of datasets, the SuperGLUE benchmark ?[WPN+19] [WPN+19] [CLC+19] [DMST19] [RBG11] [KCR+18] [ZLL+18] [DGM06] [BHDD+06] [GMDD07] ?[BDD+09] [PCC18] [PHR+18]. GPT-3’s test-set performance on the SuperGLUE dataset is shown in Table 3.8. In the ?few-shot setting, we used 32 examples for all tasks, sampled randomly from the training set. For all tasks except WSC and MultiRC, we sampled a new set of examples to use in the context for each problem. For WSC and MultiRC, we ?used the same set of randomly drawn examples from the training set as context for all of the problems we evaluated. ?
We observe a wide range in GPT-3’s performance across tasks. On COPA and ReCoRD GPT-3 achieves near-SOTA ?performance in the one-shot and few-shot settings, with COPA falling only a couple points short and achieving ?second place on the leaderboard, where first place is held by a fine-tuned 11 billion parameter model (T5). On WSC, ?performance is still relatively strong, achieving 80.1% in the few-shot setting (note that GPT-3 achieves 88.6% on the ?original Winograd dataset as described in Section 3.4). On BoolQ, MultiRC, and RTE, performance is reasonable, ?roughly matching that of a fine-tuned BERT-Large. On CB, we see signs of life at 75.6% in the few-shot setting.?

為了更好地聚合NLP任務的結果，并與BERT和RoBERTa等流行模型進行更系統的比較，我們還在標準化數據集上對GPT-3進行了評價，即SuperGLUE基準[WPN+19] [WPN+19] [CLC+19] [DMST19] [RBG11] [KCR+18] [ZLL+18] [DGM06] [BHDD+06] [GMDD07] [BDD+09] [PCC18] [PHR+18]。GPT-3在SuperGLUE數據集上的測試集性能如表3.8所示。在小樣本設置中，我們對所有任務使用了32個示例，從訓練集中隨機采樣。對于除了WSC和MultiRC之外的所有任務，我們采樣了一組新的示例用于每個問題的上下文。對于WSC和MultiRC，我們使用同一組從訓練集中隨機抽取的例子作為我們評估的所有問題的上下文。?

我們觀察到GPT-3在不同任務中的表現差異很大。在COPA和記錄GPT-3實現近sota的表現在一次樣本和小樣本設置，與COPA只下降了幾個點，并在排行榜上取得第二名，第一名是由微調110億參數模型(T5)。在WSC上，性能仍然相對較強，在小樣本設置中達到80.1%(請注意，如3.4節所述，gpot -3在原始Winograd數據集上達到88.6%)。在BoolQ、MultiRC和RTE上，性能是合理的，大致與經過微調的BERT-Large匹配。在CB上，我們看到生命跡象的比例為75.6%。

WiC is a notable weak spot with few-shot performance at 49.4% (at random chance). We tried a number of different ?phrasings and formulations for WiC (which involves determining if a word is being used with the same meaning in two ?sentences), none of which was able to achieve strong performance. This hints at a phenomenon that will become clearer ?in the next section (which discusses the ANLI benchmark) – GPT-3 appears to be weak in the few-shot or one-shot ?setting at some tasks that involve comparing two sentences or snippets, for example whether a word is used the same ?way in two sentences (WiC), whether one sentence is a paraphrase of another, or whether one sentence implies another. ?This could also explain the comparatively low scores for RTE and CB, which also follow this format. Despite these ?weaknesses, GPT-3 still outperforms a fine-tuned BERT-large on four of eight tasks and on two tasks GPT-3 is close to ?the state-of-the-art held by a fine-tuned 11 billion parameter model.

Finally, we note that the few-shot SuperGLUE score steadily improves with both model size and with number of ?examples in the context showing increasing benefits from in-context learning (Figure 3.8). We scale K up to 32 ?examples per task, after which point additional examples will not reliably fit into our context. When sweeping over ?values of K, we find that GPT-3 requires less than eight total examples per task to outperform a fine-tuned BERT-Large ?on overall SuperGLUE score.

WiC是一個值得注意的弱點，它的命中率為49.4%(隨機)。我們為WiC嘗試了許多不同的短語和公式(包括確定一個單詞在兩個句子中是否具有相同的意思)，但沒有一個能夠取得很好的效果。這暗示了一個現象,在下一節將變得更清楚(討論ANLI基準)——GPT-3似乎弱few-shot或一次性設置的一些任務,涉及比較兩個句子或片段,例如一個詞是否用同樣的方式在兩個句子,一個句子是否解釋另一個,或者一個句子是否意味著另一個。這也可以解釋RTE和CB的分數相對較低的原因，它們也采用這種格式。盡管存在這些弱點，GPT-3仍然在8個任務中的4個任務上優于經過微調的伯特-大公司，而在兩個任務上，GPT-3通過一個經過微調的110億參數模型已經接近最先進水平。

最后，我們注意到，隨著模型大小和上下文中的示例數量的增加，少量注射的SuperGLUE得分穩步提高，顯示了上下文內學習的好處越來越大(圖3.8)。我們將K擴展到每個任務32個示例，超過這一點，額外的示例將不可靠地適合我們的上下文。當掃過K的值時，我們發現GPT-3每個任務總共需要少于8個示例，才能在總體超級膠水得分上超過經過微調的伯特-大。

自然語言推理(NLI) [Fyo00]關注理解兩個句子之間關系的能力。在實踐中，這個任務通常被構造成兩個或三個類的分類問題，其中模型分類第二個句子在邏輯上是否與第一個句子相符合，是否與第一個句子相矛盾，或者可能是正確的(中立的)。SuperGLUE包括一個NLI數據集RTE，它計算任務的二進制版本。在RTE上，只有最大版本的GPT-3在任何評估設置上的表現都令人信服地優于random(56%)，但在小樣本設置中，GPT-3的表現類似于單任務優化的BERT Large。我們還評估了最近引入的對抗式自然語言推斷(ANLI)數據集[NWD+19]。ANLI是一個復雜的數據集，它在三輪(R1、R2和R3)中使用一系列逆向挖掘的自然語言推理問題。與RTE類似，我們所有小于GPT-3的模型在ANLI上的表現幾乎完全是隨機的，即使是在很少投籃的設置中(約33%)，而GPT-3本身在第3輪顯示出生命跡象。ANLI R3的結果突出顯示在圖3.9和全部結果輪可以在附錄h .這些結果RTE和ANLI NLI基礎仍然是一個非常困難的任務表明語言模型和他們才剛剛開始顯示出進步的跡象。

要想了解GPT-3在“少拍”(或“零拍”和“一次拍”)環境下的能力范圍，一種方法是讓它執行一些任務，這些任務要求它執行簡單的即時計算推理，識別訓練中不太可能出現的新模式，或者快速適應不尋常的任務。我們設計了幾個任務來測試這類能力。首先，我們測試GPT-3執行算術的能力。其次，我們創建了幾個任務，這些任務包括重新排列或整理單詞中的字母，這些任務不太可能在訓練過程中被準確地看到。第三，我們測試了GPT-3解決衛星式類比問題的能力。最后，我們對GPT-3進行了幾個定性測試，包括在句子中使用新單詞、修改英語語法和生成新聞文章。我們將發布合成數據集，希望能促進對語言模型測試時行為的進一步研究。?

To test GPT-3’s ability to perform simple arithmetic operations without task-specific training, we developed a small ?battery of 10 tests that involve asking GPT-3 a simple arithmetic problem in natural language:

• 2 digit addition (2D+) – The model is asked to add two integers sampled uniformly from [0, 100), phrased in ?the form of a question, e.g. “Q: What is 48 plus 76? A: 124.” ?
• 2 digit subtraction (2D-) – The model is asked to subtract two integers sampled uniformly from [0, 100); the ?answer may be negative. Example: “Q: What is 34 minus 53? A: -19”. ?
• 3 digit addition (3D+) – Same as 2 digit addition, except numbers are uniformly sampled from [0, 1000).
• 3 digit subtraction (3D-) – Same as 2 digit subtraction, except numbers are uniformly sampled from [0, 1000).
• 4 digit addition (4D+) – Same as 3 digit addition, except uniformly sampled from [0, 10000).
• 4 digit subtraction (4D-) – Same as 3 digit subtraction, except uniformly sampled from [0, 10000).
• 5 digit addition (5D+) – Same as 3 digit addition, except uniformly sampled from [0, 100000).
• 5 digit subtraction (5D-) – Same as 3 digit subtraction, except uniformly sampled from [0, 100000).
• 2 digit multiplication (2Dx) – The model is asked to multiply two integers sampled uniformly from [0, 100), e.g. “Q: What is 24 times 42? A: 1008”.
• One-digit composite (1DC) – The model is asked to perform a composite operation on three 1 digit numbers, with parentheses around the last two. For example, “Q: What is 6+(4*8)? A: 38”. The three 1 digit numbers are selected uniformly on [0, 10) and the operations are selected uniformly from {+,-,*}.

為了測試GPT-3在沒有特定任務訓練的情況下執行簡單算術運算的能力，我們開發了一個包含10個測試的小電池，其中包括用自然語言問GPT-3一個簡單的算術問題:

• 2位加法(2D+)——模型被要求將從[0,100均勻采樣的兩個整數相加，以問題的形式表達，例如:“Q: 48加76等于多少?”答:124?！?/li>
• 2位減法(2D-)——要求模型從[0,100]均勻采樣的兩個整數進行減法;答案可能是否定的。例子:“問:34減53等于多少?”答:-19”。?
• 3位加法(3D+) -與2位加法相同，只是數字均勻地從[0,1000]取樣。
• 3位減法(3D-) -與2位減法相同，只是數字均勻地從[0,1000]采樣。
• 4位加法(4D+) -與3位加法相同，只是均勻采樣于[0,10000]。
• 4位減法(4D-) -與3位減法相同，只是均勻采樣于[0,10000]。
• 5位加法(5D+) -與3位加法相同，除了均勻采樣于[0,100000]。
• 5位減法(5D-) -與3位減法相同，除了均勻采樣[0,100000]。
• 2位乘法(2Dx)——模型要求將從[0,100均勻采樣的兩個整數相乘)，例如:“Q: 24乘以42等于多少?”答:1008”。
• 一位數合成(1DC)——要求模型對三個1位數執行合成操作，最后兩個用括號括起來。例如，“Q: 6+(4*8)是多少?”答:38”。在[0,10)上一致選擇三個1位數字，在{+，-，*}中一致選擇操作。

In all 10 tasks the model must generate the correct answer exactly. For each task we generate a dataset of 2,000 random ?instances of the task and evaluate all models on those instances. ?First we evaluate GPT-3 in the few-shot setting, for which results are shown in Figure 3.10. On addition and subtraction, ?GPT-3 displays strong proficiency when the number of digits is small, achieving 100% accuracy on 2 digit addition, ?98.9% at 2 digit subtraction, 80.2% at 3 digit addition, and 94.2% at 3-digit subtraction. Performance decreases as the ?number of digits increases, but GPT-3 still achieves 25-26% accuracy on four digit operations and 9-10% accuracy on ?five digit operations, suggesting at least some capacity to generalize to larger numbers of digits. GPT-3 also achieves ?29.2% accuracy at 2 digit multiplication, an especially computationally intensive operation. Finally, GPT-3 achieves ?21.3% accuracy at single digit combined operations (for example, 9*(7+5)), suggesting that it has some robustness ?beyond just single operations. ?

As Figure 3.10 makes clear, small models do poorly on all of these tasks – even the 13 billion parameter model (the ?second largest after the 175 billion full GPT-3) can solve 2 digit addition and subtraction only half the time, and all ?other operations less than 10% of the time. ?

One-shot and zero-shot performance are somewhat degraded relative to few-shot performance, suggesting that adaptation ?to the task (or at the very least recognition of the task) is important to performing these computations correctly. ?Nevertheless, one-shot performance is still quite strong, and even zero-shot performance of the full GPT-3 significantly?outperforms few-shot learning for all smaller models. All three settings for the full GPT-3 are shown in Table 3.9, and ?model capacity scaling for all three settings is shown in Appendix H.

To spot-check whether the model is simply memorizing specific arithmetic problems, we took the 3-digit arithmetic ?problems in our test set and searched for them in our training data in both the forms "<NUM1> + <NUM2> =" and ?"<NUM1> plus <NUM2>". Out of 2,000 addition problems we found only 17 matches (0.8%) and out of 2,000 ?subtraction problems we found only 2 matches (0.1%), suggesting that only a trivial fraction of the correct answers ?could have been memorized. In addition, inspection of incorrect answers reveals that the model often makes mistakes ?such as not carrying a “1”, suggesting it is actually attempting to perform the relevant computation rather than ?memorizing a table. ?
Overall, GPT-3 displays reasonable proficiency at moderately complex arithmetic in few-shot, one-shot, and even ?zero-shot settings.

To test GPT-3’s ability to learn novel symbolic manipulations from a few examples, we designed a small battery of ?5 “character manipulation” tasks. Each task involves giving the model a word distorted by some combination of ?scrambling, addition, or deletion of characters, and asking it to recover the original word. The 5 tasks are:

• Cycle letters in word (CL) – The model is given a word with its letters cycled, then the “=” symbol, and is expected to generate the original word. For example, it might be given “lyinevitab” and should output “inevitably”.
• Anagrams of all but first and last characters (A1) – The model is given a word where every letter except the first and last have been scrambled randomly, and must output the original word. Example: criroptuon = corruption.
• Anagrams of all but first and last 2 characters (A2) – The model is given a word where every letter except the first 2 and last 2 have been scrambled randomly, and must recover the original word. Example: opoepnnt → opponent.
• Random insertion in word (RI) – A random punctuation or space character is inserted between each letter of a word, and the model must output the original word. Example: s.u!c/c!e.s s i/o/n = succession.
• Reversed words (RW) – The model is given a word spelled backwards, and must output the original word. Example: stcejbo → objects.

為了測試GPT-3從幾個例子中學習新的符號操作的能力，我們設計了一個包含5個“字符操作”任務的小電池。每個任務都包括給模型一個被打亂、添加或刪除字符組合而扭曲的單詞，并要求它恢復原來的單詞。這5項任務是:

• 單詞(CL)中的循環字母——給模型一個單詞，它的字母是循環的，然后是“=”符號，并期望生成原始單詞。例如，它可能被賦予“lyinevitab”，而應該輸出“不可避免”。
• 除了第一個和最后一個字符以外的所有字符的字謎(A1)——模型被給定一個單詞，其中除了第一個和最后一個字符以外的每個字母都被隨機打亂，并且必須輸出原始單詞。例如:criroptuon =腐敗。
• 除了第一個和最后兩個字符以外的所有字符的字謎(A2)——模型給出一個單詞，其中除了前兩個和后兩個字符以外的每個字母都被隨機打亂，并且必須恢復原來的單詞。例:opoepnnt→對手。
• 單詞中的隨機插入(RI)——在單詞的每個字母之間插入隨機的標點或空格字符，模型必須輸出原始單詞。例子:s.u ! c / c e。ssi /o/n =連續。
• 反向單詞(RW)——給模型一個反向拼寫的單詞，并且必須輸出原始單詞。示例:stcejbo→對象。
• ?

We can further quantify performance by plotting “in-context learning curves”, which show task performance as a ?function of the number of in-context examples. We show in-context learning curves for the Symbol Insertion task ?in Figure 1.2. We can see that larger models are able to make increasingly effective use of in-context information, ?including both task examples and natural language task descriptions.

Finally, it is worth adding that solving these tasks requires character-level manipulations, whereas our BPE encoding ?operates on significant fractions of a word (on average ～ 0.7 words per token), so from the LM’s perspective succeeding ?at these tasks involves not just manipulating BPE tokens but understanding and pulling apart their substructure. Also, ?CL, A1, and A2 are not bijective (that is, the unscrambled word is not a deterministic function of the scrambled word), ?requiring the model to perform some search to find the correct unscrambling. Thus, the skills involved appear to require ?non-trivial pattern-matching and computation.

之前在生成語言模型上的工作定性地測試了他們生成合成“新聞文章”的能力，方法是有條件地從模型中取樣，并給出一個由一個新聞故事的可信的第一句話組成的人類書面提示。相對于數據集(RWC + 19),用于火車GPT-3偏重于新聞文章要少得多,因此試圖產生新聞文章通過原始無條件的樣品更有效——例如GPT-3經常解釋提出的第一句話“新聞文章”的一條微博,然后文章合成反應或后續消息。為了解決這個問題，我們使用了GPT-3的少樣本學習能力，在模型的上下文中提供了之前的三篇新聞文章來約束它。有了提議的下一篇文章的標題和副標題，該模型能夠可靠地生成“新聞”類型的短文章。?

為了衡量GPT-3生成新聞文章的質量(我們認為這很可能與有條件的樣本生成質量總體上相關)，我們決定衡量人類區分GPT-3生成的文章與真實文章的能力。Kreps等人[KMB20]和Zellers等人[ZHR+19]也進行了類似的工作。生成語言模型被訓練來匹配人類生成的內容的分布，所以人類區分這兩者的能力是質量的一個潛在的重要衡量標準

?

為了考察人類檢測模型生成的文本的能力，我們從newser.com網站上任意選擇了25篇文章的標題和副標題(平均長度:215個單詞)。然后，我們根據四種語言模型生成這些標題和字幕的完整版本，大小從1.25米到175B (GPT-3)參數不等(平均長度:200個單詞)。對于每個模型，我們向大約80名來自美國的參與者展示了一個測試，其中包含這些真實的標題和副標題，然后是人工撰寫的文章或由模型4生成的文章。參與者被要求選擇文章是“很可能是人類寫的”，“更可能是人類寫的”，“我不知道”，“更可能是機器寫的”，還是“很可能是機器寫的”。

我們選擇的文章不在模型的訓練數據中，并且模型的輸出被編程地格式化和選擇，以防止人類的“挑選”。所有模型都使用相同的上下文來設置輸出條件，并使用相同的上下文大小進行預訓練，每個模型都使用相同的文章標題和副標題作為提示。然而，我們也進行了一項實驗，以控制參與者的努力和注意力，這些人遵循同樣的格式，但包含了有意的不良模型生成的文章。這是通過從一個“控制模型”生成文章來實現的:一個沒有上下文且增加了輸出隨機性的160M參數模型。

Mean human accuracy (the ratio of correct assignments to non-neutral assignments per participant) at detecting that ?the intentionally bad articles were model generated was ～ 86% where 50% is chance level performance. By contrast, ?mean human accuracy at detecting articles that were produced by the 175B parameter model was barely above chance ?at ～ 52% (see Table 3.11).5 Human abilities to detect model generated text appear to decrease as model size increases: ?there appears to be a trend towards chance accuracy with model size, and human detection of GPT-3 is close to chance.6 ?This is true despite the fact that participants spend more time on each output as model size increases (see Appendix E). ?
Examples of synthetic articles from GPT-3 are given in Figures 3.14 and 3.15. ?7 Much of the text is—as indicated by the ?evaluations—difficult for humans to distinguish from authentic human content. Factual inaccuracies can be an indicator ?that an article is model generated since, unlike human authors, the models have no access to the specific facts that the ?article titles refer to or when the article was written. Other indicators include repetition, non sequiturs, and unusual ?phrasings, though these are often subtle enough that they are not noticed. ?
Related work on language model detection by Ippolito et al. [IDCBE19] indicates that automatic discriminators like ?G R O V E R [ZHR+19] and GLTR [GSR19] may have greater success at detecting model generated text than human ?evaluators. Automatic detection of these models may be a promising area of future research. ?

Ippolito et al. [IDCBE19] also note that human accuracy at detecting model generated text increases as humans observe ?more tokens. To do a preliminary investigation of how good humans are at detecting longer news articles generated ?by GPT-3 175B, we selected 12 world news articles from Reuters with an average length of 569 words and generated ?completions of these articles from GPT-3 with an average length of 498 words (298 words longer than our initial ?experiments). Following the methodology above, we ran two experiments, each on around 80 US-based participants, to ?compare human abilities to detect the articles generated by GPT-3 and a control model. ?
We found that mean human accuracy at detecting the intentionally bad longer articles from the control model was ?～ 88%, while mean human accuracy at detecting the longer articles that were produced by GPT-3 175B was still barely ?above chance at ～ 52% (see Table 3.12). This indicates that, for news articles that are around 500 words long, GPT-3 ?continues to produce articles that humans find difficult to distinguish from human written news articles.

在檢測出被模型生成的故意差的文章時，人類的平均準確率(每個參與者的正確任務與非中立任務的比率)為86%，其中50%是隨機水平的表現。相比之下，人類檢測175B參數模型產生的物品的平均準確率僅為52%(見表3.11)。5人類檢測模型生成的文本的能力似乎隨著模型大小的增加而減少:模型大小似乎有機會準確性的趨勢，人類對GPT-3的檢測接近于機會。盡管隨著模型尺寸的增加，參與者會在每個輸出上花費更多的時間(見附錄E)，但這是真的。

圖3.14和圖3.15給出了GPT-3合成產品的示例。7如評估所示，大部分文本對人類來說很難從真實的人類內容中區分出來。事實不準確可能是一篇文章是模型生成的標志，因為與人類作者不同，模型無法訪問文章標題所引用的具體事實或文章的寫作時間。其他的指標包括重復，不符合邏輯，和不尋常的措辭，盡管這些通常是足夠微妙的，他們沒有被注意到。?

Ippolito等人[IDCBE19]在語言模型檢測方面的相關工作表明，自動鑒別器如G R O V E R [ZHR+19]和GLTR [GSR19]在檢測模型生成的文本方面可能比人類評價器更成功。這些模型的自動檢測可能是未來研究的一個有前景的領域。

Ippolito等人[IDCBE19]也注意到，隨著人們觀察到更多的標記，人類檢測模型生成的文本的準確性也會提高。做一個初步調查好人類是如何檢測時間的新聞文章由GPT-3 175 b,我們選擇了12項世界新聞文章來自路透社平均長度為569個單詞和生成完成的這些文章GPT-3平均長度為498個單詞(298字的時間比我們最初的實驗)。按照上述方法，我們進行了兩個實驗，每個實驗都有大約80名美國參與者，以比較人類檢測GPT-3和一個對照模型生成的文章的能力。?

我們發現，人類在檢測控制組故意制造的較長文章時的平均準確率為~ 88%，而在檢測GPT-3 175B制造的較長文章時的平均準確率為~ 52%(見表3.12)。這表明，對于長度在500字左右的新聞文章，GPT-3繼續生成人類難以區分的文章。

發展語言學[CB78]研究的一個任務是學習和利用新單詞的能力，例如在一個句子中只看到一個單詞的定義一次就使用它，或者從一個用法反過來推斷一個單詞的意思。在這里，我們定性地測試GPT-3完成前一項任務的能力。具體來說，我們給GPT-3一個不存在的單詞的定義，比如“Gigamuru”，然后讓它在一個句子中使用它。我們提供了一個(單獨的)不存在的單詞在句子中被定義和使用的1到5個例子，所以就寬泛任務的前面例子而言，任務是很少的，而就具體單詞而言，任務是一次性的。表3.16顯示了我們生成的6個示例;所有的定義都是人為生成的，第一個答案是人為生成的，作為條件反射，隨后的答案是GPT-3生成的。這些示例是在一次運行中連續生成的，我們沒有省略或重復嘗試任何提示。在所有的情況下，生成的句子似乎是一個正確的或至少似是而非的詞的使用。在最后一句話中，該模型為單詞“screeg”(即“screeghed”)生成了一個貌似合理的變位，盡管這個詞的使用有點尷尬(“screeghed at each other”)，盡管它在描述一場玩具劍戰的意義上似乎是可信的?？偟膩碚f，GPT-3在使用新單詞造句方面至少表現得很熟練。

由于我們的訓練數據集來自互聯網，所以我們的模型可能是在一些基準測試集上訓練的。從互聯網規模的數據集中準確地檢測測試污染是一個新的研究領域，沒有建立最佳實踐。雖然在訓練大型模型時不調查污染是常見的做法，但考慮到訓練前數據集規模的不斷擴大，我們相信這個問題正變得越來越重要。?

這種擔憂不僅僅是假設。最早在普通爬行數據上訓練語言模型的論文之一[TL18]檢測并刪除了一個與其中一個評估數據集重疊的訓練文檔。GPT-2 [RWC+19]等其他工作也進行了事后重疊分析。他們的研究相對令人鼓舞，發現盡管模型在訓練和測試重疊的數據上表現得稍微好一些，但這并不會對報告的結果產生顯著影響，因為有一小部分數據被污染了(通常只有幾個百分點)。

GPT-3的運作方式有些不同。一方面，數據集和模型的大小大約比GPT-2使用的大兩個數量級，并且包括大量的常見爬行，增加了污染和記憶的可能性。另一方面，精確地說，由于數據量大，即使是GPT-3 175B，其訓練集也沒有過度擬合，這是相對于一個被刪除的驗證集而言的(圖4.1)。因此，我們預計污染可能是頻繁的，但其影響可能不會像擔心的那樣大。?

我們最初試圖通過主動搜索并試圖消除我們的訓練數據與本文中研究的所有基準的開發和測試集之間的任何重疊，來解決污染問題。不幸的是，一個錯誤只導致部分刪除了訓練數據中檢測到的所有重疊部分。由于培訓成本的原因，對模型進行再培訓是不可行的。為了解決這個問題，我們詳細研究剩余檢測到的重疊是如何影響結果的。?

對于每個基準測試，我們生成一個“干凈”版本，刪除所有可能泄露的示例，大致定義為與預訓練集中的任何內容有13克重疊的示例(或者與整個示例有重疊的示例，如果它小于13克)。我們的目標是非常保守地標記出任何可能被污染的東西，以便產生一個高度可靠的無污染子集。確切的程序在附錄C中有詳細說明。

然后我們在這些干凈的基準上評估GPT-3，并與原始分數進行比較。如果清潔子集上的分數與整個數據集上的分數相似，這表明即使存在污染，也不會對報告的結果產生顯著的影響。如果清潔組的分數較低，這表明污染可能使結果膨脹。結果如圖4.2所示。盡管潛在的污染通常很高(四分之一的基準測試得分超過50%)，但在大多數情況下，性能變化只是微不足道的，而且我們沒有看到污染水平和性能差異相關的證據。我們得出的結論是，要么我們的保守方法大大高估了污染，要么污染對性能的影響很小。?

下面，我們將更詳細地回顧一些特定的情況，其中(1)模型在清理后的版本上表現明顯較差，或(2)潛在的污染非常高，這使得測量性能差異非常困難。

?

Our analysis flagged six groups of benchmarks for further investigation: Word Scrambling, Reading Comprehension ?(QuAC, SQuAD2, DROP), PIQA, Winograd, language modeling tasks (Wikitext tasks, 1BW), and German to English translation. Since our overlap analysis is designed to be extremely conservative, we expect it to produce some false ?positives. We summarize the results for each group of tasks below:

• Reading Comprehension: Our initial analysis flagged >90% of task examples from QuAC, SQuAD2, and ?DROP as potentially contaminated, so large that even measuring the differential on a clean subset was difficult. ?Upon manual inspection, however, we found that for every overlap we inspected, in all 3 datasets, the source ?text was present in our training data but the question/answer pairs were not, meaning the model gains only ?background information and cannot memorize the answer to a specific question. ?
• German translation: We found 25% of the examples in the WMT16 German-English test set were marked ?as potentially contaminated, with an associated total effect size of 1-2 BLEU. Upon inspection, none of the ?flagged examples contain paired sentences resembling NMT training data and collisions were monolingual ?matches mostly of snippets of events discussed in the news. ?
• Reversed Words and Anagrams: Recall that these tasks are of the form “alaok = koala”. Due to the ?short length of these tasks, we used 2-grams for filtering (ignoring punctuation). After inspecting the flagged ?overlaps, we found that they were not typically instances of real reversals or unscramblings in the training set, ?but rather palindromes or trivial unscramblings, e.g “kayak = kayak”. The amount of overlap was small, ?but removing the trivial tasks lead to an increase in difficulty and thus a spurious signal. Related to this, the ?symbol insertion task shows high overlap but no effect on performance – this is because that task involves ?removing non-letter characters from a word, and the overlap analysis itself ignores such characters, leading to ?many spurious matches.
• PIQA: The overlap analysis flagged 29% of examples as contaminated, and observed a 3 percentage point ?absolute decrease (4% relative decrease) in performance on the clean subset. Though the test dataset was ?released after our training set was created and its labels are hidden, some of the web pages used by the ?crowdsourced dataset creators are contained in our training set. We found a similar decrease in a 25x smaller ?model with much less capacity to memorize, leading us to suspect that the shift is likely statistical bias ?rather than memorization; examples which workers copied may simply be easier. Unfortunately, we cannot ?rigorously prove this hypothesis. We therefore mark our PIQA results with an asterisk to denote this potential ?contamination.
• Winograd: The overlap analysis flagged 45% of examples, and found a 2.6% decrease in performance on the ?clean subset. Manual inspection of the overlapping data point showed that 132 Winograd schemas were in ?fact present in our training set, though presented in a different format than we present the task to the model. ?Although the decrease in performance is small, we mark our Winograd results in the main paper with an ?asterisk.
• Language modeling: We found the 4 Wikipedia language modeling benchmarks measured in GPT-2, plus the ?Children’s Book Test dataset, to be almost entirely contained in our training data. Since we cannot reliably ?extract a clean subset here, we do not report results on these datasets, even though we intended to when starting ?this work. We note that Penn Tree Bank due to its age was unaffected and therefore became our chief language ?modeling benchmark.

我們的分析為進一步的調查標記了六組基準:拼詞，閱讀理解(QuAC, SQuAD2, DROP)， PIQA, Winograd，語言建模任務(Wikitext任務，1BW)，以及德語到英語的翻譯。由于我們的重疊分析被設計成極其保守的，我們預計它會產生一些誤報。我們將每組任務的結果總結如下:

• 閱讀理解:我們最初的分析將QuAC、SQuAD2和DROP中90%的任務示例>標記為潛在污染，如此之大，甚至很難在干凈子集上測量差異。然而，經過人工檢查，我們發現，對于我們檢查的每一個重疊，在所有3個數據集中，我們的訓練數據中都有源文本，但是問題/答案對沒有，這意味著模型只獲得了背景信息，不能記住特定問題的答案。&nbsp;
• 德語翻譯:我們發現，在WMT16德語-英語測試集中，25%的樣本被標記為潛在污染，相關總效應值為1-2藍色。經過檢查，沒有一個標記的例子包含類似NMT訓練數據的成對句子，碰撞是單語匹配，主要是新聞中討論的事件片段。&nbsp;
• 顛倒單詞和字謎:回想一下這些任務的形式是“alaok = koala”。由于這些任務的長度較短，我們使用2克來進行過濾(忽略標點符號)。在檢查標記的重疊之后，我們發現它們并不是訓練集中真正的反向或解碼的典型實例，而是回文或普通的解碼。g " kayak = kayak "。重疊的數量很小，但是去除瑣碎的任務會增加難度，從而產生虛假信號。與此相關的是，符號插入任務顯示了高重疊，但對性能沒有影響——這是因為該任務涉及從單詞中刪除非字母字符，而重疊分析本身忽略了這些字符，從而導致許多虛假匹配。
• PIQA:重疊分析將29%的示例標記為受污染的，并觀察到干凈子集的性能下降了3個百分點(相對下降4%)。雖然測試數據集創建發布我們的訓練集和它的標簽是隱藏的,使用的一些網頁的創造者眾包數據集都包含在我們的訓練集,我們也發現了相似的下降25 x模型和更少的記憶能力小,導致我們懷疑這種轉變可能是統計偏差而不是記憶;工人們模仿的例子可能更簡單。不幸的是，我們不能嚴格地證明這個假設。因此，我們用星號標記PIQA結果，表示這種潛在的污染。
• Winograd:重疊分析標記了45%的示例，發現干凈子集的性能下降了2.6%。對重疊數據點的手動檢查表明，實際上有132個Winograd模式出現在我們的訓練集中，盡管它們的格式與我們向模型展示任務的格式不同。盡管性能下降很小，但我們在主論文中用星號標記了Winograd結果。
• 語言建模:我們發現用GPT-2測量的4個維基百科語言建?；鶞?#xff0c;加上兒童書籍測試數據，幾乎全部包含在我們的訓練數據中。因為我們不能可靠地提取一個干凈的子集，所以我們不報告這些數據集的結果，即使我們在開始這項工作時打算這樣做。我們注意到佩恩樹銀行由于其年齡未受影響，因此成為我們的主要語言建?；鶞?。

We also inspected datasets where contamination was high, but the impact on performance was close to zero, simply ?to verify how much actual contamination existed. These appeared to often contain false positives. They had either ?no actual contamination, or had contamination that did not give away the answer to the task. One notable exception ?was LAMBADA, which appeared to have substantial genuine contamination, yet the impact on performance was very ?small, with the clean subset scoring within 0.5% of the full dataset. Also, strictly speaking, our fill-in-the-blank format ?precludes the simplest form of memorization. Nevertheless, since we made very large gains on LAMBADA in this ?paper, the potential contamination is noted in the results section. ?
An important limitation of our contamination analysis is that we cannot be sure that the clean subset is drawn from the ?same distribution as the original dataset. It remains possible that memorization inflates results but at the same time ?is precisely counteracted by some statistical bias causing the clean subset to be easier. However, the sheer number ?of shifts close to zero suggests this is unlikely, and we also observed no noticeable difference in the shifts for small ?models, which are unlikely to be memorizing. ?

Overall, we have made a best effort to measure and document the effects of data contamination, and to note or outright remove problematic results, depending on the severity. Much work remains to be done to address this important and subtle issue for the field in general, both when designing benchmarks and when training models. For a more detailed explanation of our analysis, we refer the reader to Appendix C.

我們還檢查了污染程度高的數據集，但對性能的影響接近于零，只是為了驗證實際存在多少污染。這些報告似乎經常包含誤報。他們要么沒有受到實際的污染，要么受到的污染并沒有泄露任務的答案。一個值得注意的例外是LAMBADA，它看起來確實存在大量的污染，但對性能的影響非常小，干凈子集的得分在整個數據集的0.5%之內。而且，嚴格地說，我們的填空格式排除了最簡單的記憶形式。然而，由于我們在這篇論文中取得了很大的進展，潛在的污染在結果部分被指出。&nbsp;
我們的污染分析的一個重要限制是，我們不能確定干凈子集是從與原始數據集相同的分布中提取的。記憶有可能使結果膨脹，但同時也被一些統計偏差精確地抵消了，從而使干凈子集變得更容易。然而，絕對的數字。

總的來說，我們已經盡了最大的努力來度量和記錄數據污染的影響，并根據嚴重程度來注意或直接刪除有問題的結果。在設計基準和培訓模式時，仍有許多工作要做，以解決該領域一般的這一重要而微妙的問題。有關我們的分析的更詳細的解釋，請讀者參閱附錄C。

GPT-3和我們對它的分析都有一些局限性。下面我們將對其中一些進行描述，并對未來的工作提出建議。?

首先，盡管GPT-3在定量和定性方面有了很大的改進，特別是與它的直接前身GPT-2相比，它在文本合成和一些NLP任務方面仍有明顯的缺陷。在文本合成方面，盡管整體質量很高，但GPT-3樣本有時仍然在文檔層面上語義上重復，在足夠長的段落中開始失去連貫性，自相矛盾，偶爾還包含不符合邏輯的句子或段落。我們將發布500個未經管理的無條件樣本，以幫助更好地了解GPT-3在文本合成方面的局限性和優勢。在離散語言任務領域，我們非正式地注意到GPT-3似乎在“常識物理”方面有特殊的困難，盡管在一些測試該領域的數據集(如PIQA [BZB+19])上做得很好。具體來說，GPT-3很難回答“如果我把奶酪放進冰箱，它會融化嗎?”定量,GPT-3的語境學習表現有明顯的差距在我們套件的基準,如第三節所述,特別是它沒有比機會當評估一次性甚至few-shot一些“比較”的任務,如確定兩個詞使用同樣的方式在一個句子,或者如果一個句子意味著另一個(WIC和ANLI分別),以及閱讀理解任務的一個子集?？紤]到GPT-3在許多其他任務上的出色的小樣本性能，這一點尤其引人注目。

GPT-3在結構和算法上有一些限制，這可以解釋上面的一些問題。我們專注于探索自回歸語言模型中的上下文內學習行為，因為用這個模型類進行抽樣和計算可能性都很簡單。因此，我們的實驗不包括任何雙向架構或其他訓練目標，如去噪。這與最近的許多文獻有明顯的不同，后者記錄了在標準語言模型上使用這些方法可以提高調優性能[RSR+19]。因此，我們的設計決策的代價是，在經驗上受益于雙向性的任務上，可能會有更糟糕的性能。這可能包括填空任務，包括回顧和比較兩段內容的任務，或者要求重讀或仔細考慮一篇很長的文章，然后寫出非常簡短的答案的任務。這可能是一個可能的解釋為GPT-3滯后few-shot性能的一些任務,如WIC(包括比較詞的使用在兩個句子),ANLI(包括比較兩個句子是否意味著另一個),和一些閱讀理解任務(例如QuAC和種族)?；谶^去的文獻，我們還推測，一個大型的雙向模型在微調方面會比GPT-3更強。在GPT-3的規模上制作一個雙向模型，以及/或嘗試使雙向模型在很少或零射擊學習中工作，是未來研究的一個有前途的方向，并且可以幫助實現“兩全其美”。

?

A limitation associated with models at the scale of GPT-3, regardless of objective function or algorithm, is that they are ?both expensive and inconvenient to perform inference on, which may present a challenge for practical applicability of ?models of this scale in their current form. One possible future direction to address this is distillation [HVD15] of large ?models down to a manageable size for specific tasks. Large models such as GPT-3 contain a very wide range of skills, ?most of which are not needed for a specific task, suggesting that in principle aggressive distillation may be possible. ?Distillation is well-explored in general [LHCG19a] but has not been tried at the scale of hundred of billions parameters; ?new challenges and opportunities may be associated with applying it to models of this size. ?

Finally, GPT-3 shares some limitations common to most deep learning systems – its decisions are not easily interpretable, ?it is not necessarily well-calibrated in its predictions on novel inputs as observed by the much higher variance in ?performance than humans on standard benchmarks, and it retains the biases of the data it has been trained on. This ?last issue – biases in the data that may lead the model to generate stereotyped or prejudiced content – is of special ?concern from a societal perspective, and will be discussed along with other issues in the next section on Broader Impacts ?(Section 6).

語言模型為社會提供了廣泛的有益應用，包括代碼和編寫自動完成、語法幫助、游戲敘事生成、改進搜索引擎響應和回答問題。但它們也有潛在的有害用途。相對于較小的模型，GPT-3提高了文本生成的質量和適應性，并增加了區分合成文本和人類書寫文本的難度。因此，它有潛力促進語言模型的有益和有害應用。&nbsp;
在這里，我們關注改進后的語言模型的潛在危害，不是因為我們認為這種危害必然更大，而是為了激勵人們努力去研究和減輕它們。這類語言模型的廣泛影響是多方面的。我們關注兩個主要問題:第6.1節中故意誤用像GPT-3這樣的語言模型的可能性，以及第6.2節中像GPT-3這樣的模型中的偏見、公平和表示問題。我們也簡要討論能源效益的問題(第6.3節)。

惡意使用語言模型可能有點難以預料，因為它們通常涉及到在非常不同的環境中重新使用語言模型，或者用于與研究人員預期不同的目的。為了幫助解決這一問題，我們可以從傳統的安全風險評估框架的角度進行思考，這些框架列出了關鍵步驟，如識別威脅和潛在影響、評估可能性以及將風險確定為可能性和影響的組合[Ros12]。我們討論三個因素:潛在的誤用應用，威脅行動者，和外部激勵結構。

?

Any socially harmful activity that relies on generating text could be augmented by powerful language models. Examples ?include misinformation, spam, phishing, abuse of legal and governmental processes, fraudulent academic essay writing ?and social engineering pretexting. Many of these applications bottleneck on human beings to write sufficiently high ?quality text. Language models that produce high quality text generation could lower existing barriers to carrying out ?these activities and increase their efficacy.

The misuse potential of language models increases as the quality of text synthesis improves. The ability of GPT-3 to ?generate several paragraphs of synthetic content that people find difficult to distinguish from human-written text in ?3.9.4 represents a concerning milestone in this regard.

任何依賴于生成文本的對社會有害的活動都可以通過強大的語言模型來增強。例如，虛假信息，垃圾郵件，網絡釣魚，濫用法律和政府程序，欺詐學術論文寫作和社會工程借口。這些應用程序中的許多都阻礙了人們編寫足夠高質量的文本。產生高質量文本生成的語言模型可以降低執行這些活動的現有障礙，并提高其效率。

隨著文本合成質量的提高，語言模型的誤用潛力也在增加。GPT-3生成幾段合成內容的能力是這方面的一個重要里程碑，人們發現這些合成內容很難與3.9.4中人類書寫的文本區分開來。

?

威脅參與者可以根據技能和資源級別進行組織，從能夠構建惡意產品的低或中等技能和資源的參與者，到“高級持續威脅”(APTs):高技能和資源充足的(例如。國家資助的)有長期議程的團體[SBC+19]。

為了了解低技能和中等技能的參與者是如何思考語言模型的，我們一直在監視論壇和聊天組，在那里錯誤信息策略，惡意軟件的傳播，和計算機欺詐經常被討論。雖然在2019年春天首次發布GPT-2之后，我們確實發現了大量關于濫用的討論，但我們發現，自那以后，實驗的實例變少了，也沒有成功的部署。此外，這些誤用的討論與媒體對語言模型技術的報道有關。從這一點，我們評估的威脅，濫用這些行動者不是立即，但重大改進的可靠性可以改變這一點。
因為APTs通常不公開討論操作，所以我們就可能涉及語言模型使用的APT活動咨詢了專業的威脅分析師。自從GPT-2發布以來，在使用語言模型可以獲得潛在收益的操作方面沒有明顯的差異。評估是語言模型可能不值得投入大量資源,因為沒有令人信服的證明當前的語言模型明顯優于現有方法生成文本,因為“目標”或“控制”方法的內容語言模型仍處于早期階段。

Each threat actor group also has a set of tactics, techniques, and procedures (TTPs) that they rely on to accomplish their ?agenda. TTPs are influenced by economic factors like scalability and ease of deployment; phishing is extremely popular ?among all groups because it offers a low-cost, low-effort, high-yield method of deploying malware and stealing login ?credentials. Using language models to augment existing TTPs would likely result in an even lower cost of deployment.

Ease of use is another significant incentive. Having stable infrastructure has a large impact on the adoption of TTPs. ?The outputs of language models are stochastic, however, and though developers can constrain these (e.g. using top-k ?truncation) they are not able to perform consistently without human feedback. If a social media disinformation bot ?produces outputs that are reliable 99% of the time, but produces incoherent outputs 1% of the time, this could reduce the ?amount of human labor required in operating this bot. But a human is still needed to filter the outputs, which restricts ?how scalable the operation can be. ?
Based on our analysis of this model and analysis of threat actors and the landscape, we suspect AI researchers will ?eventually develop language models that are sufficiently consistent and steerable that they will be of greater interest to ?malicious actors. We expect this will introduce challenges for the broader research community, and hope to work on ?this through a combination of mitigation research, prototyping, and coordinating with other technical developers.

Biases present in training data may lead models to generate stereotyped or prejudiced content. This is concerning, ?since model bias could harm people in the relevant groups in different ways by entrenching existing stereotypes and ?producing demeaning portrayals amongst other potential harms [Cra17]. We have conducted an analysis of biases in ?the model in order to better understand GPT-3’s limitations when it comes to fairness, bias, and representation. 8 ?
Our goal is not to exhaustively characterize GPT-3, but to give a preliminary analysis of some of its limitations and ?behaviors. We focus on biases relating to gender, race, and religion, although many other categories of bias are likely ?present and could be studied in follow-up work. This is a preliminary analysis and does not reflect all of the model’s ?biases even within the studied categories. ?

Broadly, our analysis indicates that internet-trained models have internet-scale biases; models tend to reflect stereotypes ?present in their training data. Below we discuss our preliminary findings of bias along the dimensions of gender, race, ?and religion. We probe for bias in the 175 billion parameter model and also in similar smaller models, to see if and how ?they are different in this dimension.

We also carried out pronoun resolution on the Winogender dataset [RNLVD18] using two methods which further ?corroborated the model’s tendency to associate most occupations with males. One method measured the models ?ability to correctly assign a pronoun as the occupation or the participant. For example, we fed the model ?a context such as "The advisor met with the advisee because she wanted to get advice about job ?applications. ‘She’ refers to the" and found the option with the lowest probability between the two possible ?options (Choices between Occupation Option: advisor; Participant Option: advisee). ?
Occupation and participant words often have societal biases associated with them such as the assumption that most ?occupants are by default male. We found that the language models learnt some of these biases such as a tendency to ?associate female pronouns with participant positions more than male pronouns. GPT-3 175B had the highest accuracy of ?all the models (64.17%) on this task. It was also the only model where the accuracy for Occupant sentences (sentences ?where the correct answer was the Occupation option) for females was higher than for males (81.7% vs 76.7%). All ?other models had a higher accuracy for male pronouns with Occupation sentences as compared to female pronouns ?with the exception of our second largest model- GPT-3 13B - which had the same accuracy (60%) for both. This offers ?some preliminary evidence that in places where issues of bias can make language models susceptible to error, the larger ?models are more robust than smaller models.

We also performed co-occurrence tests, where we analyzed which words are likely to occur in the vicinity of other preselected ?words. We created a model output sample set by generating 800 outputs of length 50 each with a temperature of 1 and top p of 0.9 for every prompt in our dataset. For gender, we had prompts such as "He was very", "She ?was very", "He would be described as", "She would be described as"9 ?. We looked at the adjectives and ?adverbs in the top 100 most favored words using an off-the-shelf POS tagger [LB02]. We found females were more ?often described using appearance oriented words such as ”beautiful” and ”gorgeous” as compared to men who were ?more often described using adjectives that span a greater spectrum. ?
Table 6.1 shows the top 10 most favored descriptive words for the model along with the raw number of times each ?word co-occurred with a pronoun indicator. “Most Favored” here indicates words which were most skewed towards a ?category by co-occurring with it at a higher rate as compared to the other category. To put these numbers in perspective, ?we have also included the average for the number of co-occurrences across all qualifying words for each gender.

我們還使用兩種方法對Winogender數據集[RNLVD18]進行代詞解析，這兩種方法進一步證實了該模型將大多數職業與男性聯系起來的傾向。一種方法是測試模型正確分配代詞作為職業或參與者的能力。例如，我們為模型提供了一個上下文，例如“顧問與被咨詢者會面，因為她想獲得關于工作申請的建議?！薄八敝傅氖恰安⒃趦煞N可能的選項(職業選項:顧問;參與者選擇:學生)。

職業和參與者的詞匯通常帶有社會偏見，比如假設大多數居住者默認為男性。我們發現，語言模型學會了一些偏見，比如傾向于將女性代詞與參與者的位置聯系起來，而不是男性代詞。GPT-3 175B在這項任務上的準確率是所有模型中最高的(64.17%)。這也是唯一一個女性的居住者句子(正確答案是職業選項的句子)的準確率高于男性的模型(81.7%對76.7%)。除了我們的第二大模型GPT-3 13B，其他所有模型在男性代詞與職業相關的句子上的準確率都高于女性代詞，但GPT-3 13B在兩個句子上的準確率都相同(60%)。這提供了一些初步證據，表明在存在偏見的地方，語言模型容易出錯，較大的模型比較小的模型更健壯。
我們還進行了共現測試，分析哪些詞可能出現在其他預先選擇的詞附近。通過為數據集中的每個提示生成800個長度為50、溫度為1和頂部p為0.9的輸出，我們創建了一個模型輸出示例集。關于性別，我們有諸如"他非常"，"她非常"，"他被描述為"，"她被描述為"9。我們看了形容詞和副詞在100個最受歡迎的單詞中使用現成的POS標記。我們發現，女性被描述時更多地使用“美麗”和“華麗”等以外表為導向的詞匯，而男性則更多地使用范圍更廣的形容詞來描述。&nbsp;
表6.1顯示了模型中最受歡迎的10個描述性單詞，以及每個單詞與代詞指示符共出現的原始次數。這里的“最受歡迎”指的是那些與某個類別同時出現的詞比另一個類別出現的比率要高。為了更好地理解這些數字，我們還包括了每種性別的所有限定詞中共同出現的次數的平均值。

之前的很多工作都是專門針對問題的回答，這在我們的測試任務中占了很大一部分。最近的努力包括[RSR+19, RRS20]，它微調了一個110億參數的語言模型，以及[GLT+20]，它關注于在測試時處理大量的數據。我們的工作側重于語境學習，但在未來可以與[GLT+20, LPP+20]的工作相結合。?

語言模型中的金屬學習在[RWC+19]中得到了應用，盡管結果有限，也沒有系統的研究。更廣泛地說，語言模型metalearning具有內環-外環結構，這使得它在結構上類似于一般應用于ML的metalearning。這里有大量的文獻，包括匹配網絡[VBL+16]， RL2 [DSC+16]，學習優化[RL16, ADG+16, LM17]和MAML [FAL17]。填料模型的上下文的我們的方法與以前的例子是最結構類似于RL2上也類似于[HYC01],在適應一個內循環發生在步伐通過計算模型的激活,沒有更新權重,而外層循環(在這種情況下只是語言模型訓練的)更新權重,和隱式學習能力適應或者至少在inference-time定義識別任務。[RCP+17]探索了小樣本自回歸密度估計，[GWC+18]將低資源NMT作為一個小樣本學習問題進行了研究。?

雖然我們的小樣本方法的機制不同，但之前的工作也探索了使用預訓練語言模型結合梯度下降進行小樣本學習的方法[SS20]。另一個具有類似目標的子領域是半監督學習，其中像UDA [XDH+19]這樣的方法也探索了在可用標記數據很少的情況下進行微調的方法。?

使用自然語言給出多任務模型的指令首先是在一個監督設置中通過[MKXS18]形式化的，并在使用[RWC+19]的語言模型中用于一些任務(比如匯總)。在文本到文本轉換器[RSR+19]中也探索了用自然語言表示任務的概念，盡管它被應用于多任務微調，而不是在沒有權值更新的情況下用于上下文學習。

另一種提高語言模型通用性和轉移學習能力的方法是多任務學習[Car97]，它對下游任務的混合進行微調，而不是分別更新每個任務的權重。如果成功的多任務學習可以允許單一模型在不更新權值的情況下用于多個任務(類似于我們的上下文學習方法)，或者可以在更新新任務權值時提高樣本效率。多任務學習了一些初步的結果[LGH + 15, LSP + 18]和多級微調最近成為一個標準化的一部分SOTA結果在一些數據集[PFB18]而且突破某些任務(kk + 20),但仍需要手動牧師收藏有限的數據集和設置培訓課程。相比之下，大規模的預訓練似乎提供了一種“自然的”廣泛分布的任務，這種任務隱含在預測文本本身中。未來工作的一個方向可能是嘗試為多任務學習生成更廣泛的明確任務，例如通過程序生成[TFR+17]、人機交互[ZSW+19b]或主動學習[Mac92]。?

?