--- license: apache-2.0 base_model: google/flan-t5-large tags: - generated_from_trainer - NLPPaper_to_Question_Generation - Summarization - Long Document Summarization model-index: - name: FLAN-T5-NLP-Paper-to-Question-Generation results: [] widget: - text: >- Generate Question, Answer pair correspond to the following research paper. [Abstract] The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data. [Introduction] Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [38, 24, 15]. Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states ht, as a function of the previous hidden state ht−1 and the input for position t. This inherently sequential nature precludes parallelization within training examples, which becomes critical at longer sequence lengths, as memory constraints limit batching across examples. Recent work has achieved significant improvements in computational efficiency through factorization tricks [21] and conditional computation [32], while also improving model performance in case of the latter. The fundamental constraint of sequential computation, however, remains. Attention mechanisms have become an integral part of compelling sequence modeling and transduction models in various tasks, allowing modeling of dependencies without regard to their distance in the input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms are used in conjunction with a recurrent network. In this work we propose the Transformer, a model architecture eschewing recurrence and instead relying entirely on an attention mechanism to draw global dependencies between input and output. The Transformer allows for significantly more parallelization and can reach a new state of the art in translation quality after being trained for as little as twelve hours on eight P100 GPUs. Question, Answer: example_title: Attention Is All You Need - text: >- Generate Question, Answer pair correspond to the following research paper. [Abstract] In this work, we explore prompt tuning, a simple yet effective mechanism for learning soft prompts to condition frozen language models to perform specific downstream tasks. Unlike the discrete text prompts used by GPT-3, soft prompts are learned through backpropagation and can be tuned to incorporate signal from any number of labeled examples. Our end-to-end learned approach outperforms GPT-3's few-shot learning by a large margin. More remarkably, through ablations on model size using T5, we show that prompt tuning becomes more competitive with scale: as models exceed billions of parameters, our method closes the gap and matches the strong performance of model tuning (where all model weights are tuned). This finding is especially relevant in that large models are costly to share and serve, and the ability to reuse one frozen model for multiple downstream tasks can ease this burden. Our method can be seen as a simplification of the recently proposed prefix tuning of Li and Liang (2021), and we provide a comparison to this and other similar approaches. Finally, we show that conditioning a frozen model with soft prompts confers benefits in robustness to domain transfer, as compared to full model tuning. [Introduction] With the wide success of pre-trained large language models, a range of techniques has arisen to adapt these general-purpose models to downstream tasks. ELMo (Peters et al., 2018) proposed freezing the pre-trained model and learning a task-specific weighting of its per-layer representations. However, since GPT (Radford et al., 2018) and BERT (Devlin et al., 2019), the dominant adaptation technique has been model tuning (or fine-tuning), where all model parameters are tuned during adaptation, as proposed by Howard and Ruder (2018).More recently, Brown et al. (2020) showed that prompt design (or priming) is surprisingly effective at modulating a frozen GPT-3 model’s behavior through text prompts. Prompts are typically composed of a task description and/or several canonical examples. This return to freezing pre-trained models is appealing, especially as model size continues to increase. Rather than requiring a separate copy of the model for each downstream task, a single generalist model can simultaneously serve many different tasks. Unfortunately, prompt-based adaptation has several key drawbacks. Task description is error-prone and requires human involvement, and the effectiveness of a prompt is limited by how much conditioning text can fit into the model’s input. As a result, downstream task quality still lags far behind that of tuned models. For instance, GPT-3 175B fewshot performance on SuperGLUE is 17.5 points below fine-tuned T5-XXL (Raffel et al., 2020) (71.8 vs. 89.3) despite using 16 times more parameters. Several efforts to automate prompt design have been recently proposed. Shin et al. (2020) propose a search algorithm over the discrete space of words, guided by the downstream application training data. While this technique outperforms manual prompt design, there is still a gap relative to model tuning. Li and Liang (2021) propose prefix tuning and show strong results on generative tasks. This method freezes the model parameters and backpropagates the error during tuning to prefix activations prepended to each layer in the encoder stack, including the input layer. Hambardzumyan et al. (2021) simplify this recipe by restricting the trainable parameters to the input and output subnetworks of a masked language model, and show reasonable results on classifications tasks. In this paper, we propose prompt tuning as a further simplification for adapting language models. We freeze the entire pre-trained model and only allow an additional k tunable tokens per downstream task to be prepended to the input text. This soft prompt is trained end-to-end and can condense the signal from a full labeled dataset, allowing our method to outperform few-shot prompts and close the quality gap with model tuning (Figure 1). At the same time, since a single pre-trained model is recycled for all downstream tasks, we retain the efficient serving benefits of frozen models (Figure 2). While we developed our method concurrently with Li and Liang (2021) and Hambardzumyan et al. (2021), we are the first to show that prompt tuning alone (with no intermediate-layer prefixes or task-specific output layers) is sufficient to be competitive with model tuning. Through detailed experiments in sections 2–3, we demonstrate that language model capacity is a key ingredient for these approaches to succeed. As Figure 1 shows, prompt tuning becomes more competitive with scale. We compare with similar approaches in Section 4. Explicitly separating task-specific parameters from the generalist parameters needed for general language-understanding has a range of additional benefits. We show in Section 5 that by capturing the task definition in the prompt while keeping the generalist parameters fixed, we are able to achieve better resilience to domain shifts. In Section 6, we show that prompt ensembling, learning multiple prompts for the same task, can boost quality and is more efficient than classic model ensembling. Finally, in Section 7, we investigate the interpretability of our learned soft prompts. In sum, our key contributions are: 1. Proposing prompt tuning and showing its competitiveness with model tuning in the regime of large language models. 2. Ablating many design choices, and showing quality and robustness improve with scale. 3. Showing prompt tuning outperforms model tuning on domain shift problems. 4. Proposing prompt ensembling and showing its effectiveness. Question, Answer: example_title: PEFT (2104.08691) - text: >- Generate Question, Answer pair correspond to the following research paper. [Abstract] For the first time in the world, we succeeded in synthesizing the room-temperature superconductor (Tc≥400 K, 127∘C) working at ambient pressure with a modified lead-apatite (LK-99) structure. The superconductivity of LK-99 is proved with the Critical temperature (Tc), Zero-resistivity, Critical current (Ic), Critical magnetic field (Hc), and the Meissner effect. The superconductivity of LK-99 originates from minute structural distortion by a slight volume shrinkage (0.48 %), not by external factors such as temperature and pressure. The shrinkage is caused by Cu2+ substitution of Pb2+(2) ions in the insulating network of Pb(2)-phosphate and it generates the stress. It concurrently transfers to Pb(1) of the cylindrical column resulting in distortion of the cylindrical column interface, which creates superconducting quantum wells (SQWs) in the interface. The heat capacity results indicated that the new model is suitable for explaining the superconductivity of LK-99. The unique structure of LK-99 that allows the minute distorted structure to be maintained in the interfaces is the most important factor that LK-99 maintains and exhibits superconductivity at room temperatures and ambient pressure. [Introduction] Since the discovery of the first superconductor(1), many efforts to search for new roomtemperature superconductors have been carried out worldwide(2, 3) through their experimental clarity or/and theoretical perspectives(4-8). The recent success of developing room-temperature superconductors with hydrogen sulfide(9) and yttrium super-hydride(10) has great attention worldwide, which is expected by strong electron-phonon coupling theory with high-frequency hydrogen phonon modes(11, 12). However, it is difficult to apply them to actual application devices in daily life because of the tremendously high pressure, and more efforts are being made to overcome the high-pressure problem(13). For the first time in the world, we report the success in synthesizing a room-temperature and ambient-pressure superconductor with a chemical approach to solve the temperature and pressure problem. We named the first room temperature and ambient pressure superconductor LK-99. The superconductivity of LK-99 proved with the Critical temperature (Tc), Zero-resistivity, Critical current (Ic), Critical magnetic field (Hc), and Meissner effect(14, 15). Several data were collected and analyzed in detail to figure out the puzzle of superconductivity of LK-99: X-ray diffraction (XRD), X-ray photoelectron spectroscopy (XPS), Electron Paramagnetic Resonance Spectroscopy (EPR), Heat Capacity, and Superconducting quantum interference device (SQUID) data. Henceforth in this paper, we will report and discuss our new findings including superconducting quantum wells associated with the superconductivity of LK-99. Question, Answer: example_title: LK-99 (Not NLP) - text: >- Generate Question, Answer pair correspond to the following research paper. [Abstract] Abstract Evaluation practices in natural language generation (NLG) have many known flaws, but improved evaluation approaches are rarely widely adopted. This issue has become more urgent, since neural NLG models have improved to the point where they can often no longer be distinguished based on the surfacelevel features that older metrics rely on. This paper surveys the issues with human and automatic model evaluations and with commonly used datasets in NLG that have been pointed out over the past 20 years. We summarize, categorize, and discuss how researchers have been addressing these issues and what their findings mean for the current state of model evaluations. Building on those insights, we lay out a long-term vision for NLG evaluation and propose concrete steps for researchers to improve their evaluation processes. Finally, we analyze 66 NLG papers from recent NLP conferences in how well they already follow these suggestions and identify which areas require more drastic changes to the status quo. [Introduction] There are many issues with the evaluation of models that generate natural language. For example, datasets are often constructed in a way that prevents measuring tail effects of robustness, and they almost exclusively cover English. Most automated metrics measure only similarity between model output and references instead of fine-grained quality aspects (and even that poorly). Human evaluations have a high variance and, due to insufficient documentation, rarely produce replicable results. These issues have become more urgent as the nature of models that generate language has changed without significant changes to how they are being evaluated. While evaluation methods can capture surface-level improvements in text generated by state-of-the-art models (such as increased fluency) to some extent, they are ill-suited to detect issues with the content of model outputs, for example if they are not attributable to input information. These ineffective evaluations lead to overestimates of model capabilities. Deeper analyses uncover that popular models fail even at simple tasks by taking shortcuts, overfitting, hallucinating, and not being in accordance with their communicative goals. Identifying these shortcomings, many recent papers critique evaluation techniques or propose new ones. But almost none of the suggestions are followed or new techniques used. There is an incentive mismatch between conducting high-quality evaluations and publishing new models or modeling techniques. While general-purpose evaluation techniques could lower the barrier of entry for incorporating evaluation advances into model development, their development requires resources that are hard to come by, including model outputs on validation and test sets or large quantities of human assessments of such outputs. Moreover, some issues, like the refinement of datasets, require iterative processes where many researchers collaborate. All this leads to a circular dependency where evaluations of generation models can be improved only if generation models use better evaluations. We find that there is a systemic difference between selecting the best model and characterizing how good this model really is. Current evaluation techniques focus on the first, while the second is required to detect crucial issues. More emphasis needs to be put on measuring and reporting model limitations, rather than focusing on producing the highest performance numbers. To that end, this paper surveys analyses and critiques of evaluation approaches (sections 3 and 4) and of commonly used NLG datasets (section 5). Drawing on their insights, we describe how researchers developing modeling techniques can help to improve and subsequently benefit from better evaluations with methods available today (section 6). Expanding on existing work on model documentation and formal evaluation processes (Mitchell et al., 2019; Ribeiro et al., 2020), we propose releasing evaluation reports which focus on demonstrating NLG model shortcomings using evaluation suites. These reports should apply a complementary set of automatic metrics, include rigorous human evaluations, and be accompanied by data releases that allow for re-analysis with improved metrics. In an analysis of 66 recent EMNLP, INLG, and ACL papers along 29 dimensions related to our suggestions (section 7), we find that the first steps toward an improved evaluation are already frequently taken at an average rate of 27%. The analysis uncovers the dimensions that require more drastic changes in the NLG community. For example, 84% of papers already report results on multiple datasets and more than 28% point out issues in them, but we found only a single paper that contributed to the dataset documentation, leaving future researchers to re-identify those issues. We further highlight typical unsupported claims and a need for more consistent data release practices. Following the suggestions and results, we discuss how incorporating the suggestions can improve evaluation research, how the suggestions differ from similar ones made for NLU, and how better metrics can benefit model development itself (section 8). Question, Answer: example_title: NLG-Eval (2202.06935) - text: >- Generate Question, Answer pair correspond to the following research paper. [Abstract] Humans have harbored a longstanding desire to acquire additional abilities through absorption. Super Mario serves as an embodiment of this human dream, which can collect items to gain extra skills such as throwing fireballs and being temporarily invincible. In this paper, we uncover that Language Models (LMs), either encoderor decoder-based, can obtain new capabilities by assimilating the parameters of homologous models without the need for retraining or GPUs. Typically, new abilities of LMs can be imparted by Supervised Fine-Tuning (SFT), reflected in the disparity between fine-tuned and pre-trained parameters (i.e., delta parameters). We initially observe that by introducing a novel operation called DARE (Drop And REscale), most of the delta parameters can be directly set to zeros without affecting the capabilities of SFT LMs and larger models can tolerate a higher proportion of discarded parameters. Based on this observation, we further sparsify delta parameters of multiple SFT homologous models with DARE and subsequently merge them into a single model by parameter averaging. We conduct experiments on eight datasets from the GLUE benchmark with BERT and RoBERTa. We also merge WizardLM, WizardMath, and Code Alpaca based on Llama 2. Experimental results show that: (1) The delta parameter value ranges for SFT models are typically small, often within 0.005, and DARE can eliminate 99% of them effortlessly. However, once the models are continuously pre-trained, the value ranges can grow to around 0.03, making DARE impractical. We have also tried to remove fine-tuned instead of delta parameters and find that a 10% reduction can lead to drastically decreased performance (even to 0.0). This highlights that SFT merely stimulates the abilities via delta parameters rather than injecting new abilities into LMs; (2) DARE can merge multiple task-specific LMs into one LM with diverse abilities. For instance, the merger of WizardLM and WizardMath increases the GSM8K zeroshot accuracy of WizardLM from 2.2 to 66.3, retaining its instruction-following ability while surpassing WizardMath’s original 64.2 performance. All resources are available at https://github.com/yule-BUAA/MergeLM. [Introduction] Human beings have always expressed their ambition to acquire additional abilities through various ways such as movies and games. For example, in X-Men’s Apocalypse, the character can absorb the powers of other mutants to strengthen himself. Likewise, the protagonist in the Super Mario games can gain superpowers like throwing fireballs by absorbing in-game items. Large Language Models (LLMs), such as GPT-4 [45], can reasonably be considered as early iterations of artificial general intelligence systems, given their performance is remarkably close to human-level capabilities. In this paper, we astonishingly find that LMs, similar to Apocalypse and Super Mario, can enhance their capabilities by absorbing other models without the need for training or GPUs. Formally, Supervised Fine-Tuning (SFT) is the most widely adopted strategy for assigning taskspecific capabilities to LMs by optimizing their parameters [13, 67]. The effectiveness of SFT is fully evident in the alteration of the model parameters before and after SFT, referred to as delta parameters [12]. We initially demonstrate that SFT LM (either encoder- or decoder-based) always tends to acquire excessively redundant delta parameters. To be specific, we present DARE, which randomly resets some delta parameters to zeros based on a drop rate p and subsequently scales the remaining parameters by a factor of 1/(1 − p). Despite its simplicity, with the assistance of DARE, when the LM model parameters reach 70 billion, we can eliminate up to 99% delta parameters with minimal impact on model performance (see Figure 1(a)). The more parameters the LM has, the larger p it can tolerate. This discovery suggests that SFT LM indeed learns a multitude of low-rank structures akin to LoRA [25]. Thus, even when most of these structures are removed, resulting in a low-rank and extremely sparse delta parameter set, the LM can still retain its capabilities. Based on this observation, we can confidently merge multiple homologous SFT LMs (pre-trained from the same backbone) without significant concerns about the decrease in their capabilities. As long as a small portion of the delta parameters remains unaffected in the merging process, the abilities of LMs unlocked by SFT can still be preserved. We first employ DARE to eliminate redundant delta parameters in each model before merging, which can potentially mitigate the interference of parameters among multiple models [62]. Then, we apply established model merging techniques [59, 26, 44, 27, 62] to the parameters with reduced redundancy to create a single model with diverse capabilities. We conduct extensive experiments on encoder-based LMs on eight datasets from the GLUE benchmark, and decoder-based Llama 2 with three distinct abilities: instruction-following, mathematical reasoning, and code-generating. We observe that: (1) SFT LMs exhibit a substantial number of redundant delta parameters whether they are based on BERT, RoBERTa, or Llama 2. DARE allows the removal of approximately 90% or even 99% delta parameters without significantly affecting the performance of downstream tasks. The rescale operation in DARE is a crucial component to guarantee effective ablations of delta parameters. Without rescaling, removing only 10% delta parameters would noticeably affect performance. We attribute this phenomenon to the fact that rescaling helps preserve the connectivity of model parameters [46]. (2) DARE is able to enhance the performance of most existing model merging methods when merging encoder-based LMs on the eight datasets from GLUE. When it comes to larger LMs based on Llama 2, the simple parameter averaging method can already produce surprisingly good results. As shown in Figure 1(b), we merge WizardLM and WizardMath by combining DARE and parameter averaging, leading to a significant improvement of WizardLM’s mathematical reasoning ability from 2.2 to 64.2 accuracy on GSM8K, while also modestly enhancing its instruction-following ability with win rate from 67.2 to 67.5 on AlpacaEval. It is worth noticing that all these benefits are achieved by solely using CPUs without further training. Similar improvements can also be observed when merging code-generating models. (3) DARE is applicable to SFT delta parameters whose value ranges are relatively small. Different from the observations of delta parameters, dropping only 10% fine-tuned parameters would lead to a catastrophic decrease in performance, even approaching zero. We also find that the delta parameters of SFT LMs usually stay within a range of 0.005 or less, indicating minimal modifications to the pre-trained LM. However, once we continue pre-training, the delta parameters can rapidly reach around 0.03, making DARE infeasible. This further confirms that SFT primarily unlocks the abilities of the pre-trained LM, rather than introducing additional abilities. Last but not least, we have implemented an open-sourced codebase at https://github.com/ yule-BUAA/MergeLM, which integrates existing popular model merging methods and supports both encoder- and decoder-based language models. We hope this work can advance the understanding of how alignment works from the perspective of parameters. Question, Answer: example_title: LM-SuperMario (2311.03099) datasets: - UNIST-Eunchan/NLP-Paper-to-QA-Generation language: - en pipeline_tag: text2text-generation --- # FLAN-T5-NLP-Paper-to-Question-Generation This model is a fine-tuned version of [google/flan-t5-large](https://huggingface.co/google/flan-t5-large) on an [allenai/QASPER: a dataset for question answering on scientific research papers ](https://huggingface.co/datasets/allenai/qasper)-based [NLP-Paper-to-QA-Generation](https://huggingface.co/datasets/UNIST-Eunchan/NLP-Paper-to-QA-Generation) dataset. ## Target Task - NLP Paper's Abstract + Introduction --> {Question} [SEP] {Answer} - Question-based Summarization - Long Document Summarization - Scientific Paper Summarization ## (1) How to use: Inference on CPU ( Code Snippets ) - Inference can be slow on CPU ### Load model directly ```python from transformers import AutoTokenizer, AutoModelForSeq2SeqLM tokenizer = AutoTokenizer.from_pretrained("UNIST-Eunchan/FLAN-T5-NLP-Paper-to-Question-Generation") model = AutoModelForSeq2SeqLM.from_pretrained("UNIST-Eunchan/FLAN-T5-NLP-Paper-to-Question-Generation") ``` ### Prompting Input ```python txt = r""" Generate Question, Answer pair correspond to the following research paper. [Abstract] + {text['abstract']} + [Introduction] + {text['introduction']} Question, Answer: """.replace("\n", "") inputs = tokenizer(txt, max_length = 1024, truncation=True, padding="max_length", return_tensors="pt") ``` ### For Multiple Question Generation (👍) ```python num_generate_sequence = 4 #8, 16, 2, 1 summaries = model.generate(input_ids =inputs["input_ids"], max_new_tokens=100, do_sample = True, top_p = 0.95, num_return_sequences = num_generate_sequence) ``` ### For Single Question Generation ```python summaries = model.generate(input_ids =inputs["input_ids"], max_new_tokens=100, do_sample = True, top_p = 0.95) ``` ```python decoded_summaries = [tokenizer.decode(s, skip_special_tokens=False, clean_up_tokenization_spaces=True) for s in summaries] decoded_summaries = [d.replace("", " ").replace(tokenizer.pad_token, "").replace(tokenizer.eos_token, "") for d in decoded_summaries] ``` ## (2) Faster Inference on GPU - about 60x faster than (1) [CPU --> COLAB T4 GPU] ### Additional Installation ```python !pip install accelerate -q !pip install bitsandbytes -q !pip install optimum -q ``` ### Load model directly ```python import torch from transformers import AutoTokenizer, AutoModelForSeq2SeqLM,BitsAndBytesConfig from optimum.bettertransformer import BetterTransformer # load model in 4-bit quantization_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16 ) tokenizer = AutoTokenizer.from_pretrained("UNIST-Eunchan/FLAN-T5-NLP-Paper-to-Question-Generation") model = AutoModelForSeq2SeqLM.from_pretrained("UNIST-Eunchan/FLAN-T5-NLP-Paper-to-Question-Generation", quantization_config=quantization_config) model = BetterTransformer.transform(model) ``` ### For Multiple Question Generation (👍) ```python # use to(device) num_generate_sequence = 16 # (about 20 sec with Colab T4 GPU) summaries = model.generate(input_ids =inputs["input_ids"].to(device), max_new_tokens=100, do_sample = True, top_p = 0.95, num_return_sequences = num_generate_sequence) ``` ### Training results It achieves the following results on the evaluation set: - Loss: 0.4504 | Training Loss | Epoch | Step | Validation Loss | |:-------------:|:-----:|:----:|:---------------:| | No log | 0.99 | 46 | 34.6109 | | 29.7732 | 1.99 | 92 | 16.5236 | | 29.7732 | 2.98 | 138 | 4.6887 | | 7.9911 | 3.97 | 184 | 0.5679 | | 7.9911 | 4.97 | 230 | 0.4795 | | 0.6152 | 5.96 | 276 | 0.4577 | | 0.6152 | 6.95 | 322 | 0.4523 | | 0.4811 | 7.95 | 368 | 0.4509 | | 0.4811 | 8.94 | 414 | 0.4505 | | 0.4721 | 9.93 | 460 | 0.4504 | ## Model description - FLAN-T5-Large (783M) ### Generated Output Example - Our model generate 16 different Q-A Pair with top-p sampling. ```python input: r""" Generate Question, Answer pair correspond to the following research paper. [Abstract] In this work, we explore prompt tuning, a simple yet effective mechanism for learning soft prompts to condition frozen language models to perform specific downstream tasks. Unlike the discrete text prompts used by GPT-3, soft prompts are learned through backpropagation and can be tuned to incorporate signal from any number of labeled examples. Our end-to-end learned approach outperforms GPT-3's few-shot learning by a large margin. More remarkably, through ablations on model size using T5, we show that prompt tuning becomes more competitive with scale: as models exceed billions of parameters, our method closes the gap and matches the strong performance of model tuning (where all model weights are tuned). This finding is especially relevant in that large models are costly to share and serve, and the ability to reuse one frozen model for multiple downstream tasks can ease this burden. Our method can be seen as a simplification of the recently proposed prefix tuning of Li and Liang (2021), and we provide a comparison to this and other similar approaches. Finally, we show that conditioning a frozen model with soft prompts confers benefits in robustness to domain transfer, as compared to full model tuning. [Introduction] With the wide success of pre-trained large language models, a range of techniques has arisen to adapt these general-purpose models to downstream tasks. ELMo (Peters et al., 2018) proposed freezing the pre-trained model and learning a task-specific weighting of its per-layer representations. However, since GPT (Radford et al., 2018) and BERT (Devlin et al., 2019), the dominant adaptation technique has been model tuning (or fine-tuning), where all model parameters are tuned during adaptation, as proposed by Howard and Ruder (2018).More recently, Brown et al. (2020) showed that prompt design (or priming) is surprisingly effective at modulating a frozen GPT-3 model’s behavior through text prompts. Prompts are typically composed of a task description and/or several canonical examples. This return to freezing pre-trained models is appealing, especially as model size continues to increase. Rather than requiring a separate copy of the model for each downstream task, a single generalist model can simultaneously serve many different tasks. Unfortunately, prompt-based adaptation has several key drawbacks. Task description is error-prone and requires human involvement, and the effectiveness of a prompt is limited by how much conditioning text can fit into the model’s input. As a result, downstream task quality still lags far behind that of tuned models. For instance, GPT-3 175B fewshot performance on SuperGLUE is 17.5 points below fine-tuned T5-XXL (Raffel et al., 2020) (71.8 vs. 89.3) despite using 16 times more parameters. Several efforts to automate prompt design have been recently proposed. Shin et al. (2020) propose a search algorithm over the discrete space of words, guided by the downstream application training data. While this technique outperforms manual prompt design, there is still a gap relative to model tuning. Li and Liang (2021) propose prefix tuning and show strong results on generative tasks. This method freezes the model parameters and backpropagates the error during tuning to prefix activations prepended to each layer in the encoder stack, including the input layer. Hambardzumyan et al. (2021) simplify this recipe by restricting the trainable parameters to the input and output subnetworks of a masked language model, and show reasonable results on classifications tasks. In this paper, we propose prompt tuning as a further simplification for adapting language models. We freeze the entire pre-trained model and only allow an additional k tunable tokens per downstream task to be prepended to the input text. This soft prompt is trained end-to-end and can condense the signal from a full labeled dataset, allowing our method to outperform few-shot prompts and close the quality gap with model tuning (Figure 1). At the same time, since a single pre-trained model is recycled for all downstream tasks, we retain the efficient serving benefits of frozen models (Figure 2). While we developed our method concurrently with Li and Liang (2021) and Hambardzumyan et al. (2021), we are the first to show that prompt tuning alone (with no intermediate-layer prefixes or task-specific output layers) is sufficient to be competitive with model tuning. Through detailed experiments in sections 2–3, we demonstrate that language model capacity is a key ingredient for these approaches to succeed. As Figure 1 shows, prompt tuning becomes more competitive with scale. We compare with similar approaches in Section 4. Explicitly separating task-specific parameters from the generalist parameters needed for general language-understanding has a range of additional benefits. We show in Section 5 that by capturing the task definition in the prompt while keeping the generalist parameters fixed, we are able to achieve better resilience to domain shifts. In Section 6, we show that prompt ensembling, learning multiple prompts for the same task, can boost quality and is more efficient than classic model ensembling. Finally, in Section 7, we investigate the interpretability of our learned soft prompts. In sum, our key contributions are: 1. Proposing prompt tuning and showing its competitiveness with model tuning in the regime of large language models. 2. Ablating many design choices, and showing quality and robustness improve with scale. 3. Showing prompt tuning outperforms model tuning on domain shift problems. 4. Proposing prompt ensembling and showing its effectiveness. Question, Answer: """.replace("\n", "") output= [' What was the size of each untrained model?[SEP] The size of the model can be a combination of the size of all the parameters in a model', ' What are the benefits of using soft prompts?[SEP] They reduce the need to use manual prompt design and conserve machine training data', ' What is the sample size of dataset?[SEP] 22840', ' How does the method outperform some of the pre-trained models?[SEP] They successfully tune their model for two tasks, one for a few shot and the other for several downstream tasks.', ' What is the sample size of the experiments?[SEP]135 for a simple task?[SEP]32 for a more complicated task', ' What is the baseline model they tested? [SEP] GPT-3 model, with four state-of-the-art examples in a masked language model', ' What task accuracy is given by prompts?[SEP]Mixed task efficiency was 93% and accuracy 85% compared to normal noise level', ' What metrics do they use?[SEP] EMO score, VSD, and SVM scores', ' What metrics are used to assess the performance of the soft prompt training?[SEP] quality of translation, accuracy of text-to-text, robustness of domain transfer, error rate.', ' How much do they experiment with the T5 baseline?[SEP] The baseline is used for simulated benchmarks.', ' Which task are they applying their method to?[SEP]They test their approach on classifications tasks', " Why do they show that their approach outperforms GPT-3's few-shot? [SEP] This is a large project that uses a multi-task approach to train GPT-3 models. In this paper, they demonstrate that the current method outperforms both the GPT-3 few-shot and the Li and Liang prefix tuning. They also show that the prefix tuning performed much better than the model tuning. What is the difference between their experiments", ' How do they compare with other techniques? [SEP] They provide a comparison for each approach.', ' Which task is the GPT-3 model most applicable to?[SEP]Classification tasks. For which tasks does the model need a subnetwork?[SEP]Classification tasks for GPT-3', ' What is the baseline test case used for this experiment?[SEP]Pompets for a variety of tasks are trained using the same method. This is the baseline, and the baseline is used for all applications.', ' What was the size of their model?[SEP] They experimented with 0.5 m.m and 0.5 m.m respectively.'] ``` ## Inference Examples ``` If Inference API generate bad, you can use model.generate() in your code for better output! ``` - (1) Attention is All You Need - (https://arxiv.org/abs/1706.03762) - (2) The Power of Scale for Parameter-Efficient Prompt Tuning - (https://arxiv.org/abs/2104.08691) - (3)(LK-99 Paper/ Not an NLP paper) The First Room-Temperature Ambient-Pressure Superconductor - (https://arxiv.org/abs/2307.12008) - (4) Repairing the Cracked Foundation: A Survey of Obstacles in Evaluation Practices for Generated Text - (https://arxiv.org/abs/2202.06935) - (5) Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch - (https://arxiv.org/abs/2311.03099) ## Training and evaluation data - Used Dataset: [UNIST-Eunchan/NLP-Paper-to-QA-Generation](https://huggingface.co/datasets/UNIST-Eunchan/NLP-Paper-to-QA-Generation) dataset. - Train: dataset['train'] + dataset['test'] - Evaluation: dataset['validation'] ### Training hyperparameters The following hyperparameters were used during training: - learning_rate: 0.0001 - train_batch_size: 1 - eval_batch_size: 1 - seed: 42 - gradient_accumulation_steps: 16 - total_train_batch_size: 16 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lr_scheduler_type: linear - lr_scheduler_warmup_steps: 184 - num_epochs: 10