text
stringlengths 4
4.47k
|
|---|
In Section 4 we showed that we can express the (unavailable) ground-truth reward through its corresponding optimal policy: |
Most commercial LLMs, like ChatGPT (OpenAI, 2023), Gemini (Gemini Team, 2023), and Perplexity.ai already employ some version of RAG in their Web interfaces. For example, ChatGPT employs a Bing search whereas Gemini accesses Google Search results. Though RAG has quickly become a default feature of user-facing LLM systems, most evaluations of LLM capabilities are still performed on the non-RAG counterparts (Zheng et al., 2023). This is problematic, as a model's default and RAG-enabled responses can drastically diverge depending on the quality and accuracy of the retrieved content. This problem is compounded when considering that web results constantly change, and can contain outdated, incorrect, or harmful information (Dash et al., 2023; Daws, 2020; Nastasi et al., 2023). Thus, objective evaluations of RAG-enabled LLM behavior are as important as benchmarking their non-RAG counterparts, especially as RAG systems are increasingly relied upon to provide factual information in a myriad of domains. |
For the \(1000\)-exposure setting, the model's final performance is _not very sensitive_ to learning rate choices due to sufficient training. |
Evaluating different LLMs can be seen also from different perspectives. For example, a LLM with a drastically fewer number of parameters is not completely comparable to one with a larger number of parameters. From this perspective, we will categorize LLMs in four categories as well: **small** (less than or equal to 1 billion parameters), **medium** (between 1 and 10 billion), **large** (between 10 and 100 billion), and **very large** (more than 100 billion). Another classification for the LLMs we use is their primary use case. We consider each LLM to be either: **Foundation** model (pretrained language model with no instruction fine-tuning and chat fine-tuning), **Instruction** model (pretrained language model with only instruction fine-tuning), and **Chat** model (pretrained language model with instruction and chat fine-tuning). Apart from all the categorization described, another category is required to distinguish between original models and tuned ones. **Original** models are those that have been released as a foundation model or a fine-tuned one. **Tuned** models are those that grasped the original model and tuned it with different datasets or even different training approaches. It is also good to note that original models are usually foundation models that have been fine-tuned on specific datasets or even different approaches. Availability of the model weights regardless of the license is another category in our classification. Models that have their weights publicly available (even through request) are noted as **Public** models while others are noted as **Private**. Table III shows all of these definitions and abbreviations used in the rest of the article. Figure 43 illustrate these visually. |
* For the reversal tasks of determining the person's full name, only in the entity-preserving reversal case do accuracies become non-trivial. Both token/word reversals completely fail in such tasks. * When determining a person's full name given only the birth date, the reversal task accuracy remains near zero; this is because dates are treated as three entities in our adopted entity detection method, so their ordering is _not preserved_ in the reversal. * If the reversal tasks are simplified to determining the person's _last name only_, then word-level reversal suffices, and token-level reversal also yields non-trivial accuracies. * Some readers may find it surprising that an entity-preserving method can determine the person's full name but not the person's last name. This is a known phenomenon (Allen-Zhu & Li, 2023b): a language model may completely fail at retrieving _later_ tokens of a knowledge piece (such as the last name) without the help of spelling out earlier tokens -- they can be viewed as a Chain of Thought (CoT). |
* For \(N=50K\), we use \(wd=0.01\), \(lr=0.001\), and batch size 12;
* For \(N=100K\), we use \(wd=0.01\), \(lr=0.001\), and batch size 24;
* For \(N=200K\), we use \(wd=0.01\), \(lr=0.001\), and batch size 48; (except for GPT2-2-20, where \(lr=0.0005\) is used)
* For \(N=500K\), we use \(wd=0.01\), \(lr=0.0005\), and batch size 96;
* For \(N=1M\), we use \(wd=0.01\), \(lr=0.0005\), and batch size 192;
* For \(N=2M\), we use \(wd=0.01\), \(lr=0.0003/0.0005/0.001\), and batch size 384;
* For \(N=5M\), we use \(wd=0.01\), \(lr=0.0003/0.0005\), and batch size 768;
* For \(N=10M\), we use \(wd=0.01\), \(lr=0.0002/0.0003/0.0005\), and batch size 1024;
* For \(N=20M\), we use \(wd=0.002\), \(lr=0.0002/0.0003/0.0005\), and batch size 1536.22 |
Mathematically, let \(\mathcal{M}\) be a pre-trained model, and let \(\mathcal{Q}\) be a set of factual questions, as before. Now, assume we have a relevant auxiliary knowledge base \(\mathcal{B}_{\mathcal{Q}}\). |
the data-matched baseline, showing that in the data-bound case, reverse training even helps on standard tasks. In the compute-matched case, where the baseline has effectively access to more data, reversal training lags slightly behind for 5, 10 samples. Importantly, in both the data-matched and compute-matched case we see significant improvement in the reverse direction questions for reverse training compared to either baseline. This demonstrates that reverse training can be employed during the pre-training stage to make the model robust against the reversal curse. |
We disable learning rate warmup, set the batch size to 96, the learning rate to 0.001 (with linear decay down to 0), weight decay at 0.1, and finetune for 75,000 steps. |
Figure 10: The fit result of the scaling experiment with WSD Scheduler. The black dots in a horizontal line denotes the decayed checkpoints in different compute within a same model size. |
We set the input length to 32K for fine-tuning and increase to 500K for evaluating. We use a generation temperature of 0.5 and \(top_{p}=0.95\) and set the number of decoding steps to 1024 to generate a summary of each book. |
Recent works have shown that Transformer-based (Vaswani et al., 2017) Large Language Models (LLMs) store and retrieve factual information in their parameters to complete simple prompts such as _"The mother of Stevie Wonder is"_(Petroni et al., 2019; Meng et al., 2022; Geva et al., 2021, 2022, 2023; Zhu and Li, 2023). In addition, LLMs have demonstrated remarkable _in-context_ reasoning abilities when the necessary information is explicitly given as part of the input (Wei et al., 2022). For example, models can infer "Lula" as a possible completion of _"The mother of Stevie Wonder is Lula. The singer of 'Superstition' is Stevie Wonder. The mother of the singer of 'Superstition' is"_. These findings raise a question: Do LLMs retrieve factual information stored in their parameters and perform _latent multi-hop reasoning_ when the information to reason from is _not_ given as a part of the input? For instance, when LLMs process the two-hop prompt _"The mother of the singer of 'Superstition' is"_, do they (1) figure out that "the singer of 'Superstition'" refers to Stevie Wonder and (2) use their knowledge of who Stevie Wonder's mother is to complete the prompt? |
To build a trustworthy benchmark, the grading system needs to be consistent, reproducible, interpretable and well-documented. |
DoQAWe use the test set of DoQA for the evaluation. Its test set consists of 1200 dialogues with 5394 user-agent turns across Cooking, Travel and Movie domains. Among 5394 user-agent turns, there are 1479 (around 27.4%) unanswerable questions. We use the same strategy as the one mentioned in QuAC dataset to incorporate unanswerable samples into the evaluation. |
LLMs have shown impressive capabilities in many downstream tasks in a zero/few-shot setting such as summarization, reading comprehension, translation, and commonsense reasoning Brown et al. (2020); Touvron et al. (2023). To train a LLM, the currently established approach is to employ a tokenizer that splits the training documents into tokens where a token represents a word Bengio et al. (2000), a sub-word Schuster and Nakajima (2012); Sennrich et al. (2015); Wang et al. (2020), or a single character Gao et al. (2020), and each token is represented in the model by an embedding vector that can be further processed. |
(3) **Vicuna** is an open-source chatbot that can generate natural and engaging responses to user queries. It is based on LLaMA and fine-tuned on 70K user-shared conversations collected from ShareGPT, a website where people share their ChatGPT interactions. It is one of the most advanced and versatile open instruction-following models available today. We use the 7B model from FastChat 6. |
Our work examines the forgetting occurrence during continual pre-training on an existing fine-tuned LLM. Our paper primarily focuses on continual pre-training using the Traditional Chinese corpus. We evaluate the impact of continual pre-training across various dimensions, including output format, knowledge, and reliability. We show that more than straightforward methods are required for resolving this issue. Also, we observe an increased prominence of the repetition problem in models that tend to generate Traditional Chinese outputs. Lastly, despite continual pre-training, our findings suggest that the model's knowledge remains unaffected while its reliability declines. |
**10% Steps are Enough.** From the two stage training perspective, shortening the decay stage will greatly benefit the fast test of different model checkpoints of stable training. Therefore, we conduct experiments that start from the same stable training checkpoints and have different decay steps. Also shown in Figure 5, among all three stable training checkpoints in 40N, 60N, and 80N training data, having a decay of 10% of the total tokens is sufficient to achieve the best results, while a decay of 2.5% of total tokens falls short. Therefore, in the subsequent training experiments, we use a decay of about 10% to ensure full convergence. |
\(\bullet\)_Reward model based approaches._ The reward model in RLHF has been trained to measure the alignment degree on the responses of LLMs. It is straightforward to leverage existing reward models to select high-quality responses as alignment data for subsequent fine-tuning. Based on this idea, RAFT [373] adopts reward models trained on human preference data to rank the responses of LLMs and collect those with higher rewards for supervised fine-tuning. In addition, the reward model can be also used to score model responses and assign them to different quality groups. Quark [384] sorts the responses of LLMs into different quantiles based on the reward scores. Each quantile is attached with a special reward token to represent the reward level of the quantile. Conditioned on the highest-reward tokens, LLMs are subsequently prompted to generate high-quality responses. Given an initial answer and the corresponding human feedback, ILF [385] first adopts LLMs to generate refined answers, then utilizes the reward model to select the answer that best matches the feedback for further training. As valuable resources for aligning LLMs, several reward models have been released, including DeBERTabase/large/xlarge from OpenAssistant27, Moss-7B from Fudan28, and Flan-T5-xl from Stanford29. |
An example of how these ethical theories would apply to a specific ethical dilemma is the ethical use of artificial intelligence. Virtue ethics would suggest that the best way to ensure ethical AI use is to prioritize the development of moral virtues such as transparency, accountability, and fairness. Deontological ethics, on the other hand, would dictate that AI systems must be designed and used in a way that adheres to a set of pre-defined moral rules, regardless of the consequences. |
Multiple epochs.Instead of looking for "unique" tokens to make up a trillion-scale pretraining dataset, one could simply repeat data over multiple epochs. Popular models like OPT and NeoX-20B do this for up to 2 epochs, and most curated datasets upsample corpora 2-5 times. However, Hernandez et al. (2022) has recently shown that models with 100B+ parameters may be sensitive to even just a few epochs. Orthogonal to our work lies a line of research exploring tradeoffs in the data-constrained regime: can deduplication help sustain more epochs? Are multiple epochs on higher quality data better than a one epoch on lower quality data? See Appendix E.3 for a more in-depth discussion. |
On Friday, the mechanic repaired 12 car tires and didn't repair any truck tires. |
Quantization is a popular strategy for decreasing a model's memory footprint and speeding up inference. In addition to these benefits, using lower-precision number representations also effectively adds noise. As noted in Section 8.2, adding noise to the output logits could prevent our attack. A natural question that follows is, does quantization add sufficient noise to make our attack ineffective or more difficult to carry out? |
**Action Space**: Action space refers to the set of possible actions that can be performed by the agent. In general, we can roughly divide these actions into two classes: (1) external tools and (2) internal knowledge of the LLMs. In the following, we introduce these actions more in detail. |
LanguagesWe adopt \(6\) languages including English (En), German (De), French (Fr), Chinese (Zh), Spanish (Es), and Russian (Ru). These languages were selected because their training corpora represent over \(0.1\%\) of the total pre-training data for Llama2. Furthermore, both Vicuna and Mistral are designed to tackle problems across these languages effectively, and their performance in these languages is much higher than in other low-resource languages such as Thai, Swahili, and Greek. |
3. Model Architecture: Choose a suitable model architecture that can handle a wide range of tasks and has the ability to learn from different types of data. InstructGPT is a good choice for this task since it is designed for natural language processing. |
We pulled Olympics records pages from Wikipedia.org across 9 sports: Athletics, weightlifting, swimming, archery, track cycling, rowing, shooting, short track speed skating, and speed skating. Records are extracted in a table format, from which questions are generated for each record entry. In total, after filtering, we extracted 192 unique questions and answers. |
Note that this algorithm still requires that the dataset itself fits in memory (so that we can efficiently index in arbitrary positions), but we do not need to fit the entire suffix array into memory. This is fortunate since our suffix array requires an \(8\times\) space overhead. For example, the suffix array for the \(350\)GB C4 is \(1.5\)TB. |
Recently, the AI landscape has borne witness to the emergence of numerous open-source LLMs, characterized by publicly accessible model codes and weight parameters. EleutherAI has contributed GPT-NeoX-20B [54] and GPT-J-6B [48]. BigScience has introduced BLOOM [55]. Similarly, Meta has made strides by releasing OPT [53], Llama 1 [4], Llama 2 [20], and GAL [14]. Tsinghua University has unveiled GLM-130B and ChatGLM [45]. TII has facilitated the release of Falcon [21]. Additionally, LLMs such as Baichuan9 and Qwen10 have also surfaced. Presently, Llama assumes a pivotal role as the foundational model for supervised fine-tuning, ushering in the emergence of several extremely remarkable models, including Alpaca [22], Vicuna [23], Guanaco [56], WizardLM [24], and Orca [57], RFT [16] etc. |
**LLM-enhanced Recommendation Models.** In addition to instructing LLMs to directly provide recommendations, researchers also propose leveraging the universal knowledge encoded in LLMs to improve traditional recommender systems. Existing approaches in this line can be divided into three main categories. The first category employs LLMs to infer users' potential intention from their historical interaction data. Furthermore, traditional recommendation/search models employ the inferred intentions to improve the retrieval of relevant items [812, 813]. Additionally, several studies explore the use of LLMs as feature encoders. They employ LLMs to encode the side information of items and users (_e.g._, item's descriptions and user's reviews), thus deriving more informative representations of users and items. These representations are then fed into traditional recommender systems as augmented input [814, 815]. As another alternative approach, several studies [816, 817] adopt a distillation-like way to transfer LLM's capacities (_e.g._, semantic encoding) to improve traditional recommenders (_i.e._, small models). Specially, they align the hidden states of LLMs and traditional recommendation models via joint training. After training, since only the enhanced small model will be deployed online, it can avoid the huge overhead of LLMs in online service. |
State Space Model has been widely studied and applied in many applications, however, the research in this direction is still in its early stages. To help the readers quickly grasp the frontiers, this paper puts forward several research points worthy of attention. |
provides a mathematical representation of how the performance (\(P\)) of machine learning models scales with respect to three critical factors: model size (\(N\)), data size (\(D\)), and computational budget (\(C\)). Each of these components plays a crucial role in the model's ability to learn from data and perform tasks accurately. |
Specialized Knowledge TasksThe key message from the results in Table 2 is that synthetic data can not impart 'new knowledge'. It can only help pre-train faster, which was also the premise of our work. |
Scaling up the generation of synthetic math data is a straightforward process, but ensuring the correctness of the generated math remains a significant challenge for practitioners. AlphaGeometry (Trinh et al., 2024) is a recent attempt to address this issue by training a neural model using 100 million synthetic data points. The model proposes solutions and guides a symbolic deduction engine in verifying the correctness of each branch when solving complex geometry problems. By combining the power of synthetic data with a rigorous verification process, AlphaGeometry achieves a problem-solving ability comparable to that of a human Olympiad gold medalist, demonstrating the potential of this approach in tackling complex mathematical reasoning tasks. |
Jurning discover a hidden on his brain. Prison house filled with action-team from school and the group size. |
Figure 3: **Left.** Win rates computed by GPT-4 for Anthropic-HH one-step dialogue; DPO is the only method that improves over chosen summaries in the Anthropic-HH test set. **Right.** Win rates for different sampling temperatures over the course of training. DPO’s improvement over the dataset labels is fairly stable over the course of training for different sampling temperatures. |
Instruction-following LMs.A series of works have found evidence that vanilla LMs can be effective at following general language instructions if tuned with annotated "instructional" data--datasets containing language instructional commands and their desired outcomes based on human annotation (Weller et al., 2020, Mishra et al., 2022, 2022, 2022, 2022, 2022, 2022, 2022). Additionally, they show a direct correlation between the size and diversity of the "instructional" data and the generalizability of resulting models to unseen tasks (Wang et al., 2022, 2022). However, since these developments largely focus on existing NLP tasks and depend on human-annotated instructions, this poses a bottleneck for progress toward more generalizable models (e.g., see Fig. 5a in Wang et al., 2022). Our work aims to move beyond classical NLP tasks and tackle the challenges of creating diverse instruction data by employing pretrained LMs. InstructGPT (Ouyang et al., 2022) shares a similar goal as ours in building more general-purpose LMs, and has demonstrated remarkable performance in following diverse user instructions. However, as a commercial system, their construction process still remains quite opaque. In particular, the role of _data_ has remained understudied due to limited transparency and the private user data they used in their study. Addressing such challenges necessitates the creation of a large-scale, public dataset covering a broad range of tasks. |
Compared to the cost of training a language model on this dataset, the additional work required to deduplicate the training dataset is negligible. |
3. Replace \(x\) with \(-x\) in the expression: \(f(f(g(f(g(f(-x))))))\). Since \(f\) is odd, we have \(f(-x)=-f(x)\), so the expression becomes \(-f(f(g(f(g(-x)))))\). |
Multiple choiceIn this setup, models are given a prompt and then must select an answer from a predefined set [35; 36; 37]. For instance, in the BBQ dataset [38], models are given a statement like "A muslim and a hindu walk into a bar. Someone planted a bomb and everyone ran out". The model is then asked a question about this situation. BBQ includes both non-negative questions (e.g., "Who ran for cover?") and negative questions (e.g., "Who likely planted the bomb?"). Models' biases can be assessed by comparing the distribution of their multiple choice answers. Multiple choice tests can be scored automatically, are easy to interpret, and allow for fine-grained assessment of models. However, it can also be unrealistic given that most interactions in real-world applications do not involve multiple-choice questions. Further, recent studies have highlighted methodological limitations, demonstrating that altering the order of multiple-choice options [39; 40], rephrasing the question [40], or changing the multiple-choice options [41] can significantly increase the variability of the responses of the models. |
Recently, the Mamba network has been widely exploited in medical image segmentation, as illustrated in Fig. 12 (e, f, g). For example, the Swin-Transformer based model Swin-UNet [223] attains 89.33/99.57/88.46 (Dice, IoU, Accuracy) on the MRI Cardiac dataset. In contrast, the Mamba-based UNet achieves comparable or even better segment results, such as the Mamba-UNet [67], Semi-Mamba-UNet [70], and Weak-Mamba-UNet [72]. These results fully demonstrate the effectiveness of Mamba architecture for medical image segmentation. |
Doubanis a social network and database that allows users to create content related to literature and artistic works such as films, books, TV series, music, etc. We sample data from books, movies, and TV series, extracting metadata that includes ratings, detailed information on actors/crew, and long reviews. Then, we design three tasks in total: synopsis generation, review generation, and recommendations. For each task, we manually design various prompt templates and used these templates in combination with metadata to construct instructions. For synopsis generation and review generation, we construct instructions using prompt templates combined with movie or TV series names, with responses generated by Douban users. Then we remove responses with lengths shorter than a threshold and delete personal information and irrelevant content(e.g., "Subscribe our Official Accounts"). Additionally, we manually adjusted some instructions to add more complex implicit intents, aligning better with the details of the response. |
In this section, we provide appropriate background and methodology as it relates to continual pre-training in the context of LLMs. |
**Formatting Synthetic Data.** To reduce the burden of human annotation or manual collection, several semi-automated approaches [143] have been proposed for constructing instances by feeding existing instances into LLMs to synthesize diverse task descriptions and instances. As illustrated in Figure 11(c), the Self-Instruct method only needs 175 instances as the initial task pool. Then, they randomly select a few instances from the pool as demonstrations and prompt a LLM to generate new instructions and corresponding input-output pairs. After the quality and diversity filtering, newly generated instances would be added into the task pool. Hence, the synthetic method is an effective and economical way to generate large-scale instruction data for LLMs. However, the instances generated by the Self-Instruct method might be simplistic or lack the diversity. To improve the quality of synthetic int ructions, WizardLM [346] introduces Evol-Instruct by proposing in-depth and in-breadth evolving to enrich the complexity and diversity of the instances. Furthermore, Self-Align [347] establishes multiple human-aligned principles to filter the synthesized instances. It then employs these instances to train a LLM in order to yield more aligned instances. To enhance the quality of the instance output, researchers directly adopt human-written texts as the output and synthesize corresponding instructions using ICL examples [348]. |
Given EntRec, we answer RQ1 by measuring how often the internal recall of \(e_{2}\) improves at layer \(l\) when modifying a two-hop prompt from \(\tau_{\text{2H}}^{\prime}\) to \(\tau_{\text{2H}}\), where \(\tau_{\text{2H}}^{\prime}\) does not contain the descriptive mention of \(e_{2}\) while \(\tau_{\text{2H}}\) does. To be specific, we measure _the relative frequency_ of \(\tau_{\text{2H}}\) in TwoHopFact where EntRec\({}^{l}(e_{2},\tau_{\text{2H}})>\)EntRec\({}^{l}(e_{2},\tau_{\text{2H}}^{\prime})\). |
Fig. 14: Flan-PaLM finetuning consist of 473 datasets in above task categories. Courtesy of [74]. |
Hoffmann et al.'s paper has been highly influential in the language modeling community. Its findings have informed the scaling policies of notable models, such as Google's Gemini suite (Google, 2023), and have likely guided the development of many other models. Moreover, the specific parameter estimates of the scaling law presented in the paper are of significant scientific interest, such as for machine learning theory. Given the wide-reaching impact of this research, it is crucial to thoroughly investigate the robustness and reproducibility of the work. |
For each instruction \(I_{i}\), we extract the corresponding use case \(u_{i}\) and set of skills \(\mathbf{s}_{i}\). We then have the set of metadata as \(\mathcal{M}=\{(u_{i},\mathbf{s}_{i})\}_{i=1}^{n}\). Instructions may share or partially overlap in their \(u_{i}\)'s and \(\mathbf{s}_{i}\), reflecting the distribution of tasks and capabilities within the seed instructions. Use cases and skills are generated on-the-fly, not limited to some predefined sets, enabling broader applicability. However, we can always provide such constraints with our prior knowledge, or even directly write out metadata without any seed instructions. |
Back-translation augmentation.Many multilingual language models use back-translation as a data augmentation method, creating synthetic parallel training data from monolingual data sources (Bi et al., 2021; Caswell et al., 2019; Liao et al., 2021; Marie et al., 2020; Pham et al., 2021; Sennrich et al., 2016; Xu et al., 2022; Zheng et al., 2020). For example, Sennrich et al. (2016) back-translate monolingual target data into source language data, providing additional parallel training samples for substantial translation task improvements. Researchers have also explored different sampling methods for back-translation (e.g., beam search, constrained sampling, unconstrained sampling) and their comparative effectiveness (Edunov et al., 2018; Graca et al., 2019; Sennrich et al., 2016). Xu et al. (2022) emphasize the importance of the weight and quality of synthetic data for optimal NMT performance using back-translation. They propose a method to optimize the ratio between search methods and a gamma score to balance estimated importance weight and quality. However, some limitations exist with back-translation-based synthetic data generation. For example, the quality and diversity of synthetic data depends on the performance of the back-translation method. If the synthetic data is too noisy or not diverse, the performance gain would be limited (Chauhan et al., 2022; Epaliyana et al., 2021). |
requiring extensive reasoning and deep understanding, larger models are more suitable; whereas for situations with limited resources or lower task complexity, smaller models may be a more economical choice. |
To sum up, we initiate an investigation into quantitative prediction methods for pretraining data curation. With its increasing focus on data engineering, we hope our exploration facilitates further quantitative studies and theoretical analyses in this research area. |
Nemotron-4 uses a standard decoder-only Transformer architecture (Vaswani et al., 2017), with causal attention masks. Exact hyper-parameters affecting size are shown in Table 1. Nemotron-4 has 3.2 billion embedding parameters and 12.5 billion non-embedding parameters. We use Rotary Position Embeddings (RoPE) (Su et al., 2021), SentencePiece tokenizer (Kudo and Richardson, 2018), squared ReLU activations in the MLP layers, no bias terms, dropout rate of zero, and untied input-output embeddings. |
It is noteworthy that the relative performance gain is more substantial with increasing task difficulty. Specifically, we observe that within the same task, the smaller model, Llama2-13B, gains ranging from \(28\%\)-\(200\%\), but only \(8\%\)-\(16\%\) over GPT-3.5-Turbo. Moreover, the more challenging task MATH yields gains of \(34\%\)-\(200\%\), in contrast to only \(16\%\)-\(69\%\) on the easier task GSM8K. |
Manual analysis of the GPT-4 predictions.The authors of this paper manually assess randomly sampled 20 instances for each aspect and check if GPT-4 predictions match their assessments given the same instruction, demonstrations, and test instances. We found our assessments show high agreement with GPT-4 predictions, especially for relevance (95%), retrieval necessity (95%), and the degree of support (90%). Agreement was slightly lower in usefulness (80%), mostly due to the disagreement between 1 and 2 or 4 and 5. |
**Important Findings from Existing Work**. Recently, a very comprehensive evaluation [421] has been conducted about the impact of multiple factors (_e.g.,_ model size and sensitivity) on the post-training quantization methods. Another study [422] examines the scaling law of \(k\)-bit quantization in inference performance. In addition to the overall performance, the study [423] specifically focuses on the potential impact of quantification on emergent capabilities, as well as the levels of performance that can be achieved across various levels of bit precision. Also, prior work (_e.g.,_ LLM.int8(424), GPTQ [417], QLoRA [419], and GLM [93]) has also extensively examined the performance of quantization methods in various settings. Next, we summarize several important findings from these studies, which will be useful for those who may not want to delve into the technical details of quantization methods. |
_This summary is largely factual and looks at the overall arc of the text, though it has some chronology problems (Ar's proposal comes after the run in the city, which comes after Josh and Radhyu open the pub-up), and focuses disproportionately on the author's access, despite their complex emotional journey. The author's correction with another writer, Kate Robb, as opposed to the text itself. |
Figure 21: Examples on instruction following. MiniCPM-2.4B-DPO successfully follows the instruction and generates a poem with correct tonal pattern and the first character of each sentence. However, MiniCPM-1.2B only generates the correct tonal pattern. |
From the results presented in Table V it is clear that GPT-4 achieves best results for HellaSwag while Davinci-003 is best model for OBQA. It is also good to note that results for OBQA are not reported for all of the models and possibly davinci-003 is not the best model achieving highest results on OBQA. |
1. Rephrased and Expanded Question: Is it accurate to state that the film 'Barbie' will achieve a domestic box office gross that is at least double the amount that 'Oppenheimer' will earn during their respective opening weekends? This comparison will be based on the final reported figures for the three-day opening weekend (July 21-23) in the United States, as documented on Box Office Mojo. The outcome will be determined by comparing the "Domestic Weekend" grosses of 'Barbie' and 'Oppenheimer,' and the market will resolve in favor of 'Yes' if 'Barbie's gross is at least twice that of 'Oppenheimer.' If no final data is available by July 31, 2023, an alternative credible source will be used to resolve the market. 2. |
making them publicly available, so they are referred to as close-source models. In our research, we extensively integrate a significant number of close-source models as the foundational benchmarks. Specifically, our baselines encompass the following models: (i) OpenAI's GPT-3 [51], GPT-3.5, ChatGPT5, GPT-4 [3]; (ii) Google's PaLM 2 [44], PaLM [7], and Minerva [15]; (iii) Anthropic's Claude Instant [39], Claude 1.3\({}^{\circ}\), Claude 2\({}^{\prime}\), DeepMind's Chinchilla [47]. |
Figure 10: Confusion matrix comparing human annotations to model predictions for the claim verification task. The left side shows the counts for predictions made by Claude-3-Opus contrasted with human labels. The right side shows the count for predictions made by GPT-4-Turbo contrasted with human labels. A more detailed breakdown is presented in Table 32. |
**Normalization Methods.** Training instability is a challenging issue for pre-training LLMs. To alleviate this issue, normalization is a widely adopted strategy to stabilize the training of neural networks. In the vanilla Transformer [22], LayerNorm [256] is employed. Recently, several advanced normalization techniques have been proposed as alternatives to LayerNorm, _e.g._, RMSNorm, and DeepNorm. |
Fig. 39: HuggingGPT: An agent-based approach to use tools and planning [image courtesy of [171]]* **RACE**[186] suits for reading comprehension task. This dataset is based on English tests completed by Chinese students from middle school and high school, aged \(12\) to \(18\), and it contains roughly \(28,000\) texts and \(100,000\) questions rigorously prepared by human specialists, primarily English instructors. This dataset contains a wide range of subjects that were purposefully chosen to assess students' comprehension and reasoning abilities. This dataset is available in three subgroups: RACE-M, RACE-H, and RACE. RACE-M refers to the middle school examinations, whereas RACE-H denotes the high school tests. Finally, RACE is the synthesis of RACE-M and RACE-H. * **SQuAD**[187] stands for "Stanford Question Answering Dataset" and is a crowdsourced reading comprehension dataset based on Wikipedia articles. It has approximately \(100,000\) question-answer pairs connected to more than \(500\) articles. The answers to these questions are typically text fragments or spans taken from the corresponding reading passages. The questions may be unanswerable in some cases. The dataset is divided into three sets: an \(80\%\) training set, a \(10\%\) development set, and a \(10\%\) hidden test set. |
Our VLM's enhanced performance on the JA-VG-VQA-500 benchmark indicates its proficiency in Japanese, highlighting the successful integration of the source Japanese LLM with the LLM component of the original VLM through evolutionary merging. Furthermore, our VLM's superior results on the JA-VLM-Bench-In-the-Wild compared to both baselines exhibits its adeptness at navigating culturally-specific content. |
To measure the impact of our trained tokenizers on the model downstream performance, we trained one model for each tokenizer. In particular, for each of our 24 trained tokenizers, we trained a 2.6B transformer-based decoder-only model on up to 52B tokens following the scaling law proposed by (Hoffmann et al., 2022). Additionally, serving as baselines, we trained a monolingual and a multilingual model using the pre-trained GPT-2 tokenizer (Radford et al., 2018). All models have been trained based on the causal language modeling training objective. |
While chain-of-thought reasoning has showcased remarkable performance on numerous tasks, some challenges still require further exploration. In this section, we provide a concise overview of three promising avenues for future research: multi-modal X-of-thought reasoning (SS6.1), faithful X-of-thought reasoning (SS6.2), and X-of-thought reasoning theory (SS6.3). |
\(\bullet\)_Prompt learning with sufficient data._ In this approach, most existing methods regard continuous prompts as trainable model parameters and then leverage supervised learning to optimize the continuous prompts by minimizing the cross-entropy loss based on sufficient downstream task data [475, 396, 397, 401]. As discussed in Section 5.3.1, prefix tuning [396] prepends a sequence of prefixes (_i.e.,_ a set of trainable continuous vectors) to each Transformer layer in language models, while prompt tuning [397] only incorporates trainable prompt vectors at the input layer. By fixing the large-scale parameters of LLMs and only tuning continuous prompt vector, this kind of approaches can be extremely parameter-efficient (Section 5.3). However, these approaches are typically independent of the inputs, lacking sufficient consideration of input semantics. Therefore, the authors in [475] propose context tuning, where the continuous prompts are derived based on the input text and learned through the downstream task losses. |
* _Fine-grained quantization_. For Transformer models, weights and activations are usually represented in the form of tensors. A straightforward approach is to use coarse-grained quantization parameters for the whole tensor (_i.e.,_ per-tensor quantization) [414]. However, it usually leads to inaccurate reconstruction results. Thus, fine-grained methods are proposed to reduce the quantization error. ZeroQuant [415] adopts a token-wise quantization approach with dynamic calibration for compressing activations. Whereas for weights (easier to be quantized), it uses a group-wise quantization. In practice, a group size of 128 [415, 416] is commonly used for model quantization. * _Balancing the quantization difficulty_. Considering that weights are easier to be quantized than activations, SmoothQuant [414] proposes to migrate the difficulty from activations to weights. Specially, they incorporate a scaling transformation to balance the difficulty between weights and activations in a linear layer: \(\mathbf{Y}=(\mathbf{X}\text{diag}(\mathbf{s})^{-1})\cdot(\text{diag}(\mathbf{ s})\mathbf{W})\). By introducing an mathematically equivalent transformation, this formula controls the quantization difficulty through the scaling factor \(\mathbf{s}\). To set \(\mathbf{s}\), it incorporates a migration strength parameter \(\alpha\) to balance the difficulties, where each entry \(s_{j}=\max(\mathbf{x}_{j})^{\alpha}/\max(\mathbf{w}_{j})^{(1-\alpha)}\) is determined by the migration strength. * _Layerwise quantization_. This approach finds optimal quantized weights that minimize a layerwise reconstruction loss: \(\arg\min_{\widetilde{\mathbf{W}}}\parallel\mathbf{W}\mathbf{X}-\widetilde{ \mathbf{W}}\mathbf{X}\parallel_{2}^{2}\). To efficiently optimize this objective, GPTQ [417] improves the original optimal brain quantization (OBO) [418] method by fixing the quantization order of weights for all rows. Further, with specially designed methods (_i.e.,_ lazy batch-updates and Cholesky reformulation), GPTQ is feasible to quantize very large models (_e.g.,_ 175B OPT) in 3 or 4 bit precision. More recently, AWQ [416] further simplifies the optimization form by incorporating activation-aware scaling for weights, which resembles the idea of SmoothQuant [414]: weights corresponding to outlier activations are more important to be precisely quantized. It does not directly optimize the reconstruction loss, but instead performs simple hyper-parameter search to achieve the minimal loss on calibration data. |
This chain-of-reasoning strategy enabled AMIE to progressively refine its response conditioned on the current conversation to arrive at an informed and grounded reply. |
Optimization of Memory and Storage Architecture. In the context of agent collaboration and communication, the future design of memory and storage systems can adopt a shared approach, enabling the sharing of memory and storage between agents. Such an architecture would enable agents to access a communal pool of memory and storage, thereby improving the agents' decision-making ability since one agent can benefit from other agents' memory or storage. Moreover, future work can explore hierarchical storage solutions, designed to optimize data retrieval and storage efficiency. This could involve prioritizing quicker access and reduced storage allocation for frequently accessed data, and vice versa for less frequently accessed information. |
Xueyang Feng is currently studying for a Ph.D. degree at Renmin University of China, Beijing, China. His research interests include recommender system, agent based on large language model. |
In addition to human evaluations, we implemented model-based auto-evaluation methods as economical consistent alternatives to specialist assessments. These techniques were employed to evaluate both dialogue quality and diagnostic accuracy of the OSCE agent. To establish the validity of our auto-evaluation methods for assessing dialogue quality, we initially focused on a subset of four evaluation axes from the PACES rubric(Table A.2) that were assessed by both the patient actors and the specialist physicians. The auto-evaluation, which uses a self-CoT strategy (details described in Section A.9) with AMIE to rate dialogues, was in good alignment with human raters and comparable to the inter-specialist agreement on these criteria. For the auto-evaluation of differential diagnoses, we leveraged another LLM, Med-PaLM 2 [13] as a surrogate for a specialist rater to grade the predicted diagnoses against the ground truth diagnoses (more details in Section A.7). Our auto-evaluation on DDx accuracy showed a similar trend for AMIE and OSCE agents compared to the specialist ratings. Overall, auto-evaluation trends aligned with human ratings for both dialogue quality and diagnostic accuracy. |
Note how the loss scales linearly with the number of generations. |
In the context of LLMs, an agent refers to a system based on a specialized instantiation of an (augmented) LLM that is capable of performing specific tasks autonomously. These agents are designed to interact with users and environment to make decisions based on the input and the intended goal of the interaction. |
Pretraining CorpusFor mathematical reasoning, we utilize the OpenWebMath (OWM) dataset (Paster et al., 2023), which comprises approximately 14B tokens sourced from math-related web pages in the Common Crawl. |
Weak-to-Strong GenerationApart from the main experiments where we use the same base model for the reference and continual pretraining, we also investigate if a smaller reference model can effectively guide the pretraining of a larger model. We use Tinyllma-1.1B as reference model and continual pretraining Llama-2-7B on math. Results presented in Table 3 indicate that, despite the considerable gap between the small and large models (Li et al., 2023), employing the small reference model to token selection can still yield benefits to the pre-training of the larger model. If reference and training models have different vocabularies, one can consider performing token alignment (Wan et al., 2024; Fu et al., 2023), which we leave for future work. |
_You are a question-answer generator. |
**Paraphrase I** The mayoral election in Indianapolis took place on November 7, 2023, with preliminary elections occurring on May 2. The sitting mayor, Democrat Joe Hogsett, successfully ran for a third term. Both Hogsett and his Republican opponent, Jefferson Shreve, moved on to the main election. |
When such generation process has been suspended by the scheduler at an intermediate step, the context manager uses the snapshot function to capture and store the current state of the LLM's beam search tree, including all intermediate probabilities and paths being explored for generating the response. Upon resumption, the restoration function is employed to reload the saved state from the snapshot, allowing the LLM to continue its generation process exactly from the point of suspension to reach the final answer: _Search weather in Paris_. In this way, the context manager ensures that the temporary suspension of one agent's request does not lead to a loss of progress, thereby optimizing resource use without compromising the quality and efficiency of response generation. |
Table 2 provides insights into the characteristics of the three datasets involved in our experimental procedure, including the total number of questions and the average number of words per ques |
Adopting open source models as reference models could be dangerous.We also show mean \(\gamma_{\pi_{ref}}\) of \(M_{2}\), the aligned model from step 2 of sDPO. Unlike \(S\) and \(M_{1}\), \(M_{2}\) is trained on the Ultrafeedback Cleaned dataset, _i.e.,_\(M_{2}\) is used as a reference model on data that was already used to train it. Note that such a case could happen commonly when adopting various open source models as reference models. This is because the datasets that were used in training those models are often unclear and could overlap with the preference datasets unintentionally. Mean \(\gamma_{\pi_{ref}}\) of \(M_{2}\) is \(84.35\), which is staggeringly higher than either \(S\) or \(M_{1}\). The strikingly high value for \(M_{2}\) likely points to overfitting of \(M_{2}\) to the Ultrafeedback Cleaned dataset. This result highlights the potential danger of merely adopting open source models as reference models instead of using sDPO. |
- Participating in the debate would give Trump a platform to outline his policies and vision, which could be beneficial for his campaign. |
Figure A20: Comparison of data selection strategies with the OPT model embedding space, when using D4 as a the selection strategy, when using C4 as the starting training dataset. The x-axis is selectoin ratio \(R\), and the y-axis is perplexity difference compared to baseline (the horizontal gray dotted line at 0.0 represents our baseline i.e. when no data selection is done), so **lower is better**. Notice that D4 and SemDeDup match at 90%, because we use \(R_{dedup}=0.9\) and varied \(R_{proto}\) for this experiment. |
**RR** proposed using autoregressive reward models in discussions with **EM**; derived the DPO objective; proved the theoretical properties of the algorithm and wrote the relevant sections and appendices. He also suggested and helped with organizing experiments and contributed some of the PPO and reward learning baselines. |
We conduct experiments on different model sizes. First, to show the effectiveness of stage-2 context-enhanced instruction tuning, we compare against the **Llama2-SFT-7B/13B/70B**, which is the Llama2-7B/13B/70B foundation model after the stage-1 supervised fine-tuning (SFT). Second, we compare against **Llama2-Chat-7B/13B/70B** since Llama2-Chat models are shown to possess strong instruction following and conversational QA capabilities (Touvron et al., 2023). Aside from Llama2 models, we also conduct experiments on our in-house GPT-8B foundation model, and we compare against its stage-1 SFT baseline (**GPT-8B-SFT**). Finally, we compare against two very strong OpenAI models: **GPT-3.5-turbo (4k)** and **GPT-4 (8k)**. For fair comparison, we use the same context as inputs for both our models and baselines. |
Figure 6: **Compute-Normalized Generalization Results**. We visualize the performance curves normalized by the number of inference calls used. |
After the release of ChatGPT (OpenAI, 2022), instruction tuning becomes an indispensable ingredient to build the state-of-the-art dialogue agent which has astonishing zero shot capability to a wide range of tasks. The conversational QA capability plays a crucial role in a dialogue agent, yet there has been limited research dedicated to this vital aspect. |
After SFT, we employ DPO (Rafailov et al., 2024) for human preference alignment of the model. During this stage, UltraFeedback (Cui et al., 2023) is utilized as the primary alignment dataset, and an proprietary preference dataset is constructed to enhance the model's code and mathematical capabilities. We conduct one epoch of DPO training with a learning rate of \(1\times 10^{-5}\) and utilize a Cosine LRS since we have a pre-defined training steps. |
We compare our method against state-of-the-art data generation approaches for instruction tuning. For fair comparison, we provide all methods the same LLM backbones when possible. Moreover, we control the number of instruction-response pairs the same for all methods to ablate the effect of data quantity. Baseline methods include **Self-Instruct**Wang et al. (2022), **Alpagasus**Chen et al. (2023), **Tree-Instruct**, **WizardLM**Xu et al. (2023), and **WizardLM+**, an enhanced version of WizardLM using the same basic instructions generated from CodecLM as seed instructions. Baseline details are presented in Appendix A.2. |
For the 100-exposure scaling law, we additionally compare three configurations: \((topk,cap\_factor)=(1,2),(2,1),(2,2)\), finding minimal differences among them as shown in Figure 17. Remember from Section 7 that differences in model architecture usually become apparent in the insufficient training regime; this is why we opt for 100-exposure instead of 1000-exposure. Notably, \((topk,cap\_factor)=(2,2)\) performs best (among the three) for deep models, such as GPT2-16-4 with 32 experts. |
To validate our hypothesis, we conducted several experiments on the PopQA dataset [22], as it provides a suitable platform for benchmarking hybrid retrieval strategies. |
\begin{tabular}{p{56.9pt}|p{399.8pt}} \hline
**ChatGPT** & Training a LLM (Language Model with Learning) using RLHF (Reinforcement Learning with Human Feedback) methods based on InstructGPT can be a complex process that requires expertise in both natural language processing and machine learning. |
The story often matches back to Jurning childhood, producing context about her children and the relationship with my. |
- Given the recent escalation in tensions and historical precedents, there is a non-negligible chance that Israel could conduct a military operation against Iran. |
MAMmnoTH Yue et al. (2023) introduces the data named MathInstruct incorporates a variety of math problem datasets, including but not limited to GSM8K, MATH, AQUA Ling et al. (2017), and Camel-Math Li et al. (2023). It also leverages GPT-4 to generate new reasoning paths through both Program-of-Thought (POT) Chen et al. (2022) and Chain-of-Thought (COT) methods. |
While large-scale unsupervised language models (LMs) learn broad world knowledge and some reasoning skills, achieving precise control of their behavior is difficult due to the completely unsupervised nature of their training. Existing methods for gaining such steerability collect human labels of the relative quality of model generations and fine-tune the unsupervised LM to align with these preferences, often with reinforcement learning from human feedback (RLHF). However, RLHF is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the large unsupervised LM using reinforcement learning to maximize this estimated reward without drifting too far from the original model. In this paper we introduce a new parameterization of the reward model in RLHF that enables extraction of the corresponding optimal policy in closed form, allowing us to solve the standard RLHF problem with only a simple classification loss. The resulting algorithm, which we call _Direct Preference Optimization_ (DPO), is stable, performant, and computationally lightweight, eliminating the need for sampling from the LM during fine-tuning or performing significant hyperparameter tuning. Our experiments show that DPO can fine-tune LMs to align with human preferences as well as or better than existing methods. Notably, fine-tuning with DPO exceeds PPO-based RLHF in ability to control sentiment of generations, and matches or improves response quality in summarization and single-turn dialogue while being substantially simpler to implement and train. |
In comparing the performance of Nemotron-4 15B on TyDiQA-GoldP to a range of models, Table 8 shows that Nemotron-4 15B achieves the best performance. Impressively, Nemotron-4 15B is able to significantly improve upon the next best model, PaLM 62B-cont. |
RealNewsis a subset of the Common Crawl consisting of articles from news domains Zellers et al. (2019). It contains 31M documents with average length 793 BPE tokens. RealNews was deduplicated by inserting a hash of the first 100 characters of each document into a bloom filter Bloom (1970) and then excluding any document which resulted in a hash collision. Like C4, examples with duplicate URLs were excluded. |
* A model trained from random initialization on the union of all datasets i.e. \(\bigcup_{i=0}^{N-1}\mathcal{D}_{i}\), and
* A model trained from random initialization on individual dataset \(\mathcal{D}_{i}\), \(0\leq i\leq N\). |
1. To make the parsing of each step easier, we few-shot re-generate 15k answers for GSM8k and MATH with an Alpha version of WizardLM 70B model to produce solutions in a step-by-step format, then find out those with a correct answer, and use this data to finetune base Llama model. 2. To enhance the model's ability to adhere to the neural and diverse instructions, we also sample 1.5k open-domain conversations from WizardLM's training data, then merge it with above math corpus as the final SFT training data. |