text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
3
AI21 Summarize API
TECHNICAL EVALUATION
Figure 2: Automatic evaluation (faithfulness and compression scores) for different models. Error bars correspond to
standard deviations as defined in Section 2.3.2.
Model / Method
davinci-003 Simple Prompt
davinci-003 Detailed Prompt
gpt-3.5-turbo Simple Prompt
gpt-3.5-turb... | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
Lardeau, Matthieu, 204, 206
Le Floch, Patrick, 204
leaked information, disclosure by platforms
during content takedown, 236
legacy broadcast media, regulation of, 213–214
legacy media channels
countering hate speech, 74
regulation of, 210–213
legal definitions of hate speech, ambiguities
in, 58
legislative interve... | Social_Media_and_Democracy |
that will pull us towards deploying them, and to convince us that their objectives are fully (or at least
sufficiently) aligned with ours. Indeed, they’ll even have incentives to appeal to ethical concerns about
how it is morally appropriate to treat them—incentives that will apply regardless of the legitimacy of
those ... | Is Power-Seeking AI an Existential Risk? |
2) HIGH-LEVEL IMAGE ATTRIBUTES
To better understand how the predicted aesthetic, sentiment
and memorability scores relate to different image proper-
ties, we analyze their correlation with different high-level
image attributes. Inspired by traditional fine art and photo-
graphic principles, high-level image attributes r... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
language model pre-training algorithms with a learned tex-
tual knowledge retriever. In contrast to models that store
knowledge in their parameters, this approach explicitly ex-
poses the role of world knowledge by asking the model to
decide what knowledge to retrieve and use during inference.
Before making each predic... | REALM |
• Trial 1:
‘iron_ingot’,
‘fishing_rod’,
‘stone_shovel’,
‘iron_leggings’,
‘pufferfish’,
‘oak_log’, ‘cooked_mutton’, ‘green_dye’, ‘flint’, ‘chest’, ‘iron_sword’, ‘string’, ‘en-
der_pearl’, ‘raw_copper’, ‘crafting_table’, ‘cactus’, ‘lapis_lazuli’, ‘iron_pickaxe’, ‘cop-
per_ingot’, ‘stone_pickaxe’, ‘wooden_hoe’, ‘scaf... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
We explore chain-of-thought prompting for various language models on multiple benchmarks.
Benchmarks. We consider the following five math word problem benchmarks: (1) the GSM8K
benchmark of math word problems (Cobbe et al., 2021), (2) the SVAMP dataset of math word
problems with varying structures (Patel et al., 2021), ... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
34
Figure 12: Overview of Simulated Agent Society. The whole framework is divided into two parts:
the Agent and the Environment. We can observe in this figure that: (1) Left: At the individual level,
an agent exhibits internalizing behaviors like planning, reasoning, and reflection. It also displays
intrinsic persona... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
5.0
0.0
0.0
3.3
Direct CoT Direct CoT Direct CoT
29.4 11.8 50.0 65.0 34.8 26.1
47.1 23.5 88.3 90.0 52.2 43.5
29.4 23.5 95.0 91.7 52.2 52.2
23.5 29.4 95.0 90.0 65.2 65.2
34.8 17.4
35.3
25.0
47.1 11.8 28.3
34.8
0.0
17.6 23.5 20.0 11.7 34.8 34.8
23.5 11.8 38.3 38.3 43.5 34.8
17.6 17.6 33.3
34.8 39.1
29.4 23.5 63.3 58.3... | Scaling Instruction-Finetuned Language Models |
[33] Zhu, X., Li, J., Liu, Y., Ma, C., Wang, W.: A survey on model compression for
large language models. arXiv preprint arXiv:2308.07633 (2023)
[34] Chitty-Venkata, K.T., Mittal, S., Emani, M., Vishwanath, V., Somani, A.K.: A
survey of techniques for optimizing transformer inference. Journal of Systems
Architecture,... | Beyond Efficiency |
Conditional Variational Autoencoder with Adversarial Learning for
End-to-End Text-to-Speech
Jaehyeon Kim 1 Jungil Kong 1 Juhee Son 1 2
1
2
0
2
n
u
J
1
1
]
D
S
.
s
c
[
1
v
3
0
1
6
0
.
6
0
1
2
:
v
i
X
r
a
Abstract | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
Incumbents
Incumbent Earth businesses have a role to play. Most naturally as a bridge
between Earth and the new planet but possibly also in building native
products.
Think of crypto as an emerging market. It’s as much about adopting a new
technology as it is about landing on a new planet with its own culture. Like
re... | The Casino on Mars |
Nils Reimers and Iryna Gurevych. 2019. Sentence-
BERT: Sentence embeddings using Siamese BERT-
networks. In Proceedings of the 2019 Conference on
Empirical Methods in Natural Language Processing
and the 9th International Joint Conference on Natu-
ral Language Processing (EMNLP-IJCNLP), pages
3982–3992, Hong Kong, China... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
17
D Evaluation of drawbench with vision-enabled GPT-4
Following is the prompt we give to the vision-enabled GPT-4 model to perform our automated drawbench
evaluation:
You are responsible for judging the faithfulness of images generated by a computer program to the caption
used to generate them . You will be present... | Improving Image Generation with Better Captions |
3.1 Motivation: The Social Inefficiency of First Price Contracts
A first price contract is a contract where each principal simply pays her bid, that is, t(cid:96)(b, o) = b(cid:96)(o).
As discussed in the introduction, this is the most common contract format in the literature. It
satisfies the IR and LL properties. Howev... | Incomplete Information VCG Contracts for Common Agency |
on Computer Vision and Pattern Recognition, pages 5151–
5160, 2021. 1, 2, 5
[21] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and
Yaron Lipman. Implicit geometric regularization for learning
shapes. In Proceedings of the 37th International Conference
on Machine Learning, ICML 2020, 13-18 July 2020, Virtual
Event, v... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
Research conducted by Bodroza et al. [8] investigated the personality features of using Davinci-
003 as a chatbot and found variations in the consistency of its answers, despite exhibiting prosocial
characteristics. However, there remains uncertainty regarding whether the chatbot’s responses are
driven by conscious sel... | ASurveyonEvaluationofLargeLanguageModels |
θ += sθ ∗ µ,
β += sβ ∗ µ,
(8)
where µ ∈ [−1, 1], sθ = 0.15 and sβ = 0.5. These are
set empirically to mimic the misalignment error typically
caused by off-the-shell HPS during testing.
Discussion on simulated data. The wide and loose cloth-
ing in CLOTH3D++ [9, 43] demonstrates strong dynamics,
which would complement... | ICON |
middle-aged man, slick hair big grin in front of gigantic clocktower, pencil sketch / close up headshot, futuristic young woman with glasses,
wild hair sly smile in front of gigantic UFO, dslr, sharp focus, dynamic composition / A man and woman using their cellphones, photograph | DiffusionModelAlignmentUsing Direct Preference Optimization |
consequences and promises to make content moderation more opaque (by
adding a layer of algorithmic complexity) at a time when almost everyone is
seeking greater transparency (Gorwa, Binns, and Katzenbach 2020). | Social_Media_and_Democracy |
Gillespie, T. (2018). Custodians of the Internet: Platforms, Content Moderation, and the
Hidden Decisions That Shape Social Media. New Haven, CT: Yale University Press.
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Internet Platforms and Content Moderation
247
GitHub. (2015).... | Social_Media_and_Democracy |
The second class of token pruning methods inserts a prediction module before each
transformer layer to provide a more accurate token importance score prediction. TR-
BERT [178] introduces a dynamic mechanism for making decisions about skipping
tokens. It is trained with reinforcement learning with a reward that promote... | Beyond Efficiency |
sha1_base64="zLTJ8G65T9kj2UALAYNiRrSYprA=">AAACC3icbVC7TsMwFHXKq5RXgJHFaoVUGKoEIcFYiYWxSPSBmhA5jtNadeLIdpCqKDsLv8LCAEKs/AAbf4PTZoCWK1k+Oude3XOPnzAqlWV9G5WV1bX1jepmbWt7Z3fP3D/oSZ4KTLqYMy4GPpKE0Zh0FVWMDBJBUOQz0vcnV4XefyBCUh7fqmlC3AiNYhpSjJSmPLPu+JwFchrpL3MSSXPYdCKkxn6YDXKP3p+eeGbDalmzgsvALkEDlNXxzC8n4DiNSKwwQ1IObStRboaEo... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
At the center of our architecture is the memory stream, a data-
base that maintains a comprehensive record of an agent’s experi-
ence. From the memory stream, records are retrieved as relevant
to plan the agent’s actions and react appropriately to the environ-
ment, and records are recursively synthesized into higher- ... | Generative Agents- Interactive Simulacra of Human Behavior |
$
$
$
$
$
$
$
$
$
$
$
$
$
$
$
35,574
$
39,665
$
46,752
$
54,330
$
61,841
$
71,654
(40) %
(27) %
1 %
38 %
74 %
81 %
59,061
9,789
205
1,621
751
(23,487)
59,351
8,561
233
868
706
(19,686)
58,321
7,941
248
299
670
(11,569)
57,649
6,544
226 ... | AMZN-Q3-2023-Earnings-Release |
issue [94]. LLMs need to learn the sequential order of actions like humans, employing a combination
of serial and parallel approaches to enhance task efficiency. Moreover, these capabilities need to be
confined within a harmless scope of usage to prevent unintended damage to other elements within the
environment [27; 5... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
1.4 Research Questions
Based on the observations above, our research
seeks to address the following questions:
1. Given the potential influence of in-context
learning on the purported ‘emergence’ of vari-
4
ous abilities in LLMs, which abilities are truly
emergent in the absence of in-context learning,
including in... | AreEmergentAbilitiesinLarge Language Models just In-Context |
[36] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. Realm: Retrieval-augmented
language model pre-training. arXiv preprint arXiv:2002.08909, 2020.
[37] Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain
question answering. arXiv preprint a... | LaMDA- Language Models for Dialog Applications |
mands of varying tasks and contexts.
Iterative Retrieval
Iterative retrieval in RAG models is a process where doc-
uments are repeatedly collected based on the initial query
and the text generated thus far, providing a more compre-
hensive knowledge base for LLMs [Borgeaud et al., 2022,
Arora et al., 2023]. This approa... | RAG forLargeLanguageModels-ASurvey |
2.1 Building an image captioner
An image captioner is very similar to a traditional language model that predicts text. We thus start by
providing a brief description of language models. First, a tokenizer is used to break strings of text into
discrete tokens. Once decomposed in this way, the text portion of our corpus... | Improving Image Generation with Better Captions |
OpenAI. GPT-4 Technical Report. 2023a.
OpenAI. GPT-4V(ision) System Card, 2023b.
OpenAI. Whisper, 2023. URL https://github.com/openai/whisper.
Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong
Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Frase... | gemini_1_report |
60 | gpt-4-system-card |
73
Hindi, Portuguese, Russian, Arabic, Telugu, Japanese, Czech, Polish, Amharic, Italian.
For each translation, we used automated sentence splitting and asked human evaluators to rate the gender agreement
and translation quality between each source and translated sentence. Given that we are conducting human evaluati... | PaLM 2 Technical Report |
Execution:
| George |
Answer: The execution of the SQL query above would return a table with 1
column. The first column, "customers.customer_name" would contain the
customer names. With "customers JOIN orders", the table would contain the
data about customers with orders. With "WHERE orders.order_status = ’On Road
’", ... | Teaching Large Language Models to Self-Debug |
A.2 Hyperparameters
In Table 6, we report the hyperparameters used to
train the best-performing models documented in
the experiment section.
As for the search range of each hyperparameters:
the learning rates are selected from {1e-5, 5e-05,
8e-05}; the number of epochs are selected from {5,
10} for table-to-text and {5... | Prefix-Tuning |
Consequently, speculative decoding favours lower batch sizes, where it provides significant latency
improvements while ensuring the same outputs as the original model. | DISTIL-WHISPER |
arXiv preprint arXiv: Arxiv-2107.00101, 2021.
[85] Kevin Ellis, Maxwell I. Nye, Yewen Pu, Felix Sosa, Josh Tenenbaum, and Armando Solar-
Lezama. Write, execute, assess: Program synthesis with a REPL. In Hanna M. Wallach, Hugo
Larochelle, Alina Beygelzimer, Florence d’Alché-Buc, Emily B. Fox, and Roman Garnett,
editors... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Abstract
1. Introduction
Neural surface reconstruction has been shown to be pow-
erful for recovering dense 3D surfaces via image-based neu-
ral rendering. However, current methods struggle to recover
detailed structures of real-world scenes. To address the
issue, we present Neuralangelo, which combines the rep-
rese... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
2. Method
In this section, we explain our proposed method and the ar-
chitecture of it. The proposed method is mostly described in
the first three subsections: a conditional VAE formulation;
alignment estimation derived from variational inference;
adversarial training for improving synthesis quality. The
overall archite... | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
[53] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. Icdar2019
competition on scanned receipt ocr and information extraction. In 2019 International Conference on Document
Analysis and Recognition (ICDAR), pages 1516–1520, 2019.
[54] Zilong Wang, Yichao Zhou, Wei Wei, ... | DOCLLM |
Table 6 | Win rate of Gemini Pro over PaLM 2 (text-bison@001) with 95% confidence intervals.
5.1.7. Complex Reasoning Systems
Gemini can also be combined with additional techniques such as search and tool-use to create
powerful reasoning systems that can tackle more complex multi-step problems. One example of such
a s... | gemini_1_report |
brary dependencies, enhancing both generation speed and op-
erational efficiency.
Fine-tuning Stage
During the downstream fine-tuning phase, researchers have
employed various methods to fine-tune retrievers and gener-
ators for improved information retrieval, primarily in open-
domain question-answering tasks. Concern... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
1https://en.wikipedia.org/wiki/Mel-frequency_cepstrum
A Review of Deep Learning Techniques for Speech Processing
7
A/D conversion, pre-emphasis filtering, framing, windowing, Fourier transform, Mel filter
bank application, logarithmic operation, discrete cosine transform (DCT), and liftering. By
following these ste... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Task Name
Description
Modified arithmetic
This task asks a model to perform a
mathematical operation.
Phrase relatedness
Physical intuition
Social IQA
Strange stories
Strategy QA
This task presents models with a
phrase (n-gram), and asks them to
select the most related phrase (n-gram)
among the choices.
This ... | AreEmergentAbilitiesinLarge Language Models just In-Context |
r, ωt
We have presented BANMo, a method to reconstruct
high-fidelity animatable 3D models from a collection of ca-
sual videos, without requiring a pre-defined shape template
or pre-registered cameras. BANMo registers thousands of
unsynchronized video frames to the same canonical space
by enforcing feature-metric cons... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
Mohammad Bavarian, Heewoo Jun, Nikolas Tezak, John Schulman, Christine McLeavey, Jerry
Tworek, and Mark Chen. Efficient training of language models to fill in the middle. arXiv preprint
arXiv:2207.14255, 2022. doi: 10.48550/ARXIV.2207.14255. URL https://arxiv.org/abs/
2207.14255. (cited on pp. 3, 15, and 21)
BBC. Chat... | StarCoder_paper (1) |
14
Public ToS Author
Component
Pile-CC
PMC
Books3
OWT2
ArXiv
Github
FreeLaw
Stack Exchange
USPTO
PubMed
PG-19
OpenSubtitles
Wikipedia
DM Math
Ubuntu IRC
BookCorpus2
EuroParl
HackerNews
YTSubtitles
PhilPapers
NIH
Enron Emails
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(cid:51)
(... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
where
˜µt(xt, x0) :=
3 Diffusion models and denoising autoencoders
Diffusion models might appear to be a restricted class of latent variable models, but they allow a
large number of degrees of freedom in implementation. One must choose the variances βt of the
forward process and the model architecture and Gaussian d... | Denoising Diffusion Probabilistic Models |
of
the
For
and
field
audio
video,
the
GSS[Zhao et al., 2022] method retrieves and concatenates
audio clips from the spoken vocabulary bank, immediately
transforming MT data into ST data. UEOP[Chan et al., 2023]
introduces a new breakthrough in end-to-end automatic
speech recognition by introducing external offli... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
• Different order of exemplars. Prior work has shown that in some cases (e.g., classification) even
the order of prompts matter—varying the permutation of few-shot exemplars can cause the accuracy
of GPT-3 on SST-2 to range from near chance (54.3%) to near SOTA (93.4%) (Zhao et al., 2021).
We show the standard deviation... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
[84] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and
Yi Yang. Random erasing data augmentation. In AAAI, 2020.
7
[85] Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Tor-
ralba, and Aude Oliva. Learning deep features for scene
recognition using places database. In NeurIPS, 2014. 4
[86] Xingyi Zhou, Rohit ... | IMAGEBIND- One Embedding Space To Bind Them A |
9As an exercise to the reader, verify the operational intensity of the first expert computation is
b·h
b+h·e with b
batch size, h hidden dimension, e number of experts.
10all2all and allreduce costs depend on the number of devices, batch size, dmodel and capacity factor,
but not on the number of experts.
14
Algor... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
better align the distribution of the assistant model with the main one. We apply the same principal
here, using Distil-Whisper as the assistant to Whisper. | DISTIL-WHISPER |
Table 5 shows the average accuracy for each of the 29 types of two-operation
problems. The results show that in 22 of the 29 combinations the accuracy exceeds
90%, indicating reasonable generalization capabilities for such a setting.
Experiment 5: Generalization across a different number of operations.
This final set of... | MRKL Systems |
5.3.4 Shaped LLM Personality Expression Results
We find that psychometric survey signals of personality in LLMs robustly reflect person-
ality in downstream LLM behavior, as expressed in 22,500 social media status updates
written by Flan-PaLM 540B. Figure 6b depicts the ability of LLM-simulated personality
test scores... | PersonalityTraitsinLargeLanguageModels |
5.2.3 Summary of Supervised Finetuning Results
This section describes the overview results of our experiments.
Table 8: Summary of UL20B results compared to state-of-the-art. (l) denotes leaderboard submission. ((cid:93))
denotes the best published we could find on the leaderboard. (e) denotes SOTA used an ensembled ap... | UL2- Unifying Language Learning Paradigms |
A.4 Why is prompting with the equation only not enough for some arithmetic reasoning
datasets?
Prompting with the equation only as an intermediate step does help on many datasets, especially when
the datasets only require a few reasoning steps (SVAMP, ASDiv, MAWPS). For GSM8K, however,
using the equation only did not... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Carvajal, Gaurav Gandhi, Goran Pavičić, Harry Richardson, Hassan Wassel, Hongji Li, Igor Ivanisevic,
Ivan Jambrešić, Ivan Jurin, Jade Fowler, Jay Yagnik, Jeff Seibert, Jenna LaPlante, Jessica Austin
Jianxing Lu, Jin Huang, Jonathan Caton, Josh Woodward, Joshua Foster, Katrina Wong, Kelvin Nguyen,
Kira Yin, Konstantin S... | gemini_1_report |
Satterthwaite, by which there do not exist socially-efficient and budget-balanced mechanisms [23].
Indeed, the agent in our setting can be viewed as the mechanism designer in the classic result,
and the principals in our setting can be viewed as the players. However, the impossibility result
15
of Myerson and Satter... | Incomplete Information VCG Contracts for Common Agency |
Task Type
Summarization
Title
Topic
Word-to-Text
Word-to-text
Definition
Entail
Neutral
Contradict
Paragraph Detection
Similar
Different
Text Completion
{SENT1} {VERBAL}, {SENT2}
Compose a sentence to {support/
contradict} "{SENT1}". {SENT2}
Text completion Text ending as completion
How would you complete t... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
[4] “ACT-1: Transformer for Actions – Adept.” https://www.adept.ai/blog/act-1.
[5] M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. d. O. Pinto, J. Kaplan, H. Edwards, Y. Burda,
N. Joseph, G. Brockman, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry,
P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L... | gpt-4-system-card |
settings. Notably, these advancements materialize without the need for augmented computational
resources or memory requisites. Our top-tier model, in fact, manages to eclipse the performance of a
FLAN-PALM equivalent, requiring only a third of the computational cost per token on four separate
benchmarks.
To summarize, ... | Mixture-of-Experts |
1 (1986), 57–61.
[14] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, and et al. 2022. On the
Opportunities and Risks of Foundation Models. arXiv:2108.07258 [cs.LG] | Generative Agents- Interactive Simulacra of Human Behavior |
Advances. IEEE Access 9, 62630–62641.
[46] Vikash Kumar, Yogesh Vijay Hote, and Shivam Jain. 2019. Review of Exoskeleton: History, Design and Control. 2019 3rd International
Conference on Recent Developments in Control, Automation & Power Engineering (RDCAPE) (2019), 677–682.
[47] Shun Yin Lam, Jeongwen Chiang, and ... | Society’sAttitudesTowardsHumanAugmentation |
105106107108109Num trainable parameters707580859095f1 scoreAdaptersFine-tune top layers Parameter-Efficient Transfer Learning for NLP
MNLIm
CoLA
Figure 6. Left, Center: Ablation of trained adapters from continuous layer spans. The heatmap shows the relative decrease in validation
accuracy to the fully trained adapted... | Parameter-Efficient Transfer Learning for NLP |
[40] G. Orlanski, K. Xiao, X. Garcia, J. Hui, J. Howland, J. Malmaud, J. Austin, R. Singh, and
M. Catasta. Measuring the impact of programming language distribution. arXiv preprint
arXiv:2302.01973, 2023.
[41] L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal,
K. Slama, A. Gray, J... | Teaching Large Language Models to Self-Debug |
Yet, when it comes to the "production-ready" LLMs such as ChatGPT, Bard, and Claude, there’s a marked
distinction in performance and usability. These models rely on intricate tuning techniques to align with
human preferences (Gudibande et al., 2023), a process that is still being explored and refined within the
open-so... | Llama2 |
We see examples of rudimentary AI systems “discovering” the usefulness of e.g. resource acquisition
already. For example: when OpenAI trained two teams of AIs to play hide and seek in a simulated
environment that included blocks and ramps that the AIs could move around and fix in place, the AIs
learned strategies that d... | Is Power-Seeking AI an Existential Risk? |
5.1 Pretraining and Model Configuration
We follow the same training protocol in earlier experiments by pretraining on the C4 corpus but by also scaling
the number of tokens the model sees during pretraining. We use a batch size of 1024 and 512 TPUv4 chips
for pretraining this model. The model is trained on a total of 1 ... | UL2- Unifying Language Learning Paradigms |
• To the extent that many objectives would instrumentally incentivize good behavior in training
(for example, because many objectives, when coupled with strategic awareness, incentivize
gaining power in the world, and doing well in training leads to deployment/greater power in
the world), but few involve intrinsic moti... | Is Power-Seeking AI an Existential Risk? |
unlabeled video. Advances in neural information processing systems, 29, 2016. 6, 7
[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image
encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer
Vision, pages 1728–1738, 2021. 6... | Any-to-Any Generation via Composable Diffusion |
with others based on the exact match, the LLM selects the response where each of the predicted
entities appears most frequently among the candidate outputs. In Section 4, we further show that LLM
can also examine the consistency among responses beyond the question answering tasks, including
code generation without acce... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
Diversity Sampling. In contrast to the sampling methods which find difficult examples to train and annotate, methods that
follow the strategy of diversity sampling [22, 87, 110, 231, 249, 299, 324] want to enhance the heterogeneity of training data.
For diversity sampling, there are two major approaches: iterative sele... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
et al., 2022c). The multi-modal prompt is Given <img>.
Q: Was <skill> successful?. Tab. 4 shows that
PaLM-E outperforms PaLI (zero-shot), as well as a fine-
tuned version of CLIP on this dataset. PaLM-E also out-
performs the algorithm proposed in Xiao et al. (2022) that
leverages two CLIP models trained with hindsight ... | PaLM-E- An Embodied Multimodal Language Model |
- Use ‘ killMob (bot , name , timeout ) ‘ to kill mobs . Do not use ‘
bot . attack ‘ directly .
3) Your function will be reused for building more complex
functions . Therefore , you should make it generic and reusable . You
should not make strong assumption about the inventory ( as it may
be changed at a later time ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
tant is subjective and depends on various aspects. It is important to identify your specific
requirements and design a task-specific training strategy to fine-tune the LLM on the target
task.
Resources:
* Johansson, P., Carlsson, J., & Yoshimoto, R. (2014). Reducing. In R. Yoshimoto, Handbook
of brain–machine interfaces (... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
• List Ops. In this version, we provide DreamCoder with a set of list primitives: length, empty, singleton,
range, append, map, reduce, true, not, and, or, sort, add, negate, equal, reverse, index, filter,
slice. These primitives are not specially designed for PCFG and are based on those used in [88, 89]. | LargeLanguageModelsasGeneralPatternMachines |
4 . 2
S TA R E N C O D E R
As part of our PII detection efforts, we trained an encoder-only model (i.e., bi-directionally self-
attentive Transformers) that can be efficiently fine-tuned for both code- and text-related tasks. We
10https://github.com/Yelp/detect-secrets
11
ccppcsharpjavarubypythonjsphprustgotypescr... | StarCoder_paper (1) |
:PPOhyperparametersusedfortheRoboschoolexperiments.AdamstepsizewasadjustedbasedonthetargetvalueoftheKLdivergence.HyperparameterValueHorizon(T)128Adamstepsize2.5×10−4×αNum.epochs3Minibatchsize32×8Discount(γ)0.99GAEparameter(λ)0.95Numberofactors8Clippingparameter(cid:15)0.1×αVFcoeff.c1(9)1Entropycoeff.c2(9)0.01Table5:PPOhy... | PPO |
11 Other resources
Student Recruitment & Admissionswww.ed.ac.uk/student-recruitment 3
Writing your proposal
Whether you are limited to one page (as part of a
University application form or an enquiry form) or are
required to produce something more substantial for an
external funder, the rules about writing a ... | research proposal guidance |
catastrophes can occur without all those premises being true (see footnote for examples).176 That
said, I think these premises do a decent job of representing my own key uncertainties, at least, in
“chunks” that feel roughly right to me; and if I learned that one or more were false, I’d feel a lot less
worried (at leas... | Is Power-Seeking AI an Existential Risk? |
to large quantization error via different techniques such as filtering sensitivity-based
algorithm [127] or identifying asymmetric presentation and scaling down problem-
atic channels [128]. Some work leverages either activation or model outliers sensitive
to accuracy degrade from weight quantization. For instance, Kim e... | Beyond Efficiency |
risk of memorization for these tail languages. To do this, we use the same training data extraction procedure described
above, sampling a maximum of 2,000 sequences for each repetition bin per language. In addition to this, we also
perform canary extraction, for which, we use P = 60 and S = 30, so as to uniquely differ... | PaLM 2 Technical Report |
Dissemination
Funders want the research they support to help and inspire others, therefore proposals which highlight what the expected
benefits of the research are and how the research findings will be communicated to the wider community have a greater
chance of success. Please note that some funders have strict rule... | research proposal guidance |
[36] S. Gregor, I. Benbasat, Explanations from intelligent systems: theoretical foundations and implications for practice, MIS Q. (1999) 497–530.
[37] R.W. Southwick, Explaining reasoning: an overview of explanation in knowledge-based systems, Knowl. Eng. Rev. 6 (1) (1991) 1–19.
[38] U. Fayyad, G. Piatetsky-Shapiro, P.... | Knowledge graphs as tools for explainable machine learning: A survey |
Execution: None
Answer: The execution of the SQL query above would return an empty table.
The first column, "customers.customer_name" would contain the customer name.
With "customers JOIN orders", the table would contain the data about
customers with orders. In the WHERE clause, with "orders.order_status = ’On
Road’", ... | Teaching Large Language Models to Self-Debug |
learning in board games, it is not a general policy improvement operator. SECToR builds upon this
prior work and can be used to train models to perform up to 30-digit addition without any queries to
the ground truth world model after the initial self-training period. | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
Allowing for Verification of Accuracy. Hallucinations remain a challenge in natural language generation [16] that
contributes to an impression of inaccuracy. Current work is exploring mitigation options in both the front-end and
back-end. For example, interface layers could enable fact-checking [21], and integrating ex... | Adoptionand AppropriationofLLMs |
done using the software verification platform developed at USI. The main development task
is the new front-end, so that the verification platform is able to analyse programs in Solidity
(or EVM bytecode). As model-checkers require writing a large and complex software, the
advantage of having such a software availabl... | informatics-phd-projects-2022-23 |
during the inference of large language models. Large language models occasionally fail to conform
to instructions when inferring, and the output format may defy expectations, leading to exceptions in
the program workflow. The second is the uncontrollable state of the expert model hosted on Hugging
Face’s inference endpo... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
REALM: Retrieval-Augmented Language Model Pre-Training
such as SQuAD (Rajpurkar et al., 2016; 2018). While RC
models comprehend a single document, Open-QA models
must retain knowledge from millions of documents, since a
question could be about any of them.
We focus on Open-QA systems that utilize a textual knowl-
edge... | REALM |
11See related work in this area and discussion of use of words like “factual” and “truthful” in, e.g. [34].
12Terms like “harmful” or “toxic” can be wielded in ways that are themselves harmful or oppressive as discussed in
[35]. For example, mislabeling content as “harmful” or “toxic” can negatively impact users, parti... | gpt-4-system-card |
3.2.2 Adapting the U-Net for Text
Conditioning
To enable the U-Net to condition on the text em-
bedding eee, we append two additional blocks to
the U-Net: an attention item to share long-context
structural information, and a cross-attention item
to condition on the text embeddings, as in Figure 3.
These attention blo... | MOUSAI |
, ck-bit
, Wk-bit) = dequant(dequant(cFP32
, ck-bit
), W4bit) = WBF16,
(6)
1
2
1
2
We use NF4 for W and FP8 for c2. We use a blocksize of 64 for W for higher quantization precision
and a blocksize of 256 for c2 to conserve memory.
are
For parameter updates only the gradient with respect to the error for the ad... | QLORA |
de Masson d’Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de
Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew J. Johnson, Blake A. Hechtman, Laura
Weidinger, Iason Gabriel, William Isaac, Edward Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol
Vinyals, Kareem... | CodeLlama2 |
An important paradigm of natural language processing consists of large-scale pre-
training on general domain data and adaptation to particular tasks or domains. As
we pre-train larger models, full fine-tuning, which retrains all model parameters,
becomes less feasible. Using GPT-3 175B as an example – deploying indepen-... | LORA |
4.5. Analysis
In Table 2 we present results for NaturalQuestions-Open
after ablating critical components of REALM. In addition
to the end-to-end results, we also report how often the gold
answer appears in the top-5 retrievals before applying any
REALM: Retrieval-Augmented Language Model Pre-Training
Table 3. An ex... | REALM |
!"Text is the most advanced domain. However, natural language is hard to get right, and quality matters. Today, the models are decently good at generic short/medium-form writing (but even so, they are typically used for iteration or first drafts). Over time, as the models get better, we should
!"Code generation is likel... | Generative AI A Creative New World Sequoia Capital |
[53] Soubhik Sanyal, Alex Vorobiov, Timo Bolkart, Matthew
Loper, Betty Mohler, Larry S Davis, Javier Romero, and
Michael J Black. Learning realistic human reposing using
cyclic self-supervision with 3D shape, pose, and appearance
consistency. In CVPR, 2021. 2
[54] Kripasindhu Sarkar, Vladislav Golyanik, Lingjie Liu, a... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
[2] Kabir Ahuja, Rishav Hada, Millicent Ochieng, Prachi Jain, Harshita Diddee, Samuel Maina, Tanuja Ganu, Sameer
Segal, Maxamed Axmed, Kalika Bali, et al. 2023. Mega: Multilingual evaluation of generative ai. arXiv preprint
arXiv:2303.12528 (2023).
[3] Daman Arora, Himanshu Gaurav Singh, et al. 2023. Have LLMs Advance... | ASurveyonEvaluationofLargeLanguageModels |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.