text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
that need to be addressed. Under more complex scenarios, the perceiver should be able to support multiple
modalities, such as text, vision, and audio, to capture the diverse nature of feedback from the user and the
environment. | Tool Learning with Foundation Models |
Fair Housing Council of San Fernando Valley v. Roommates.com concerned
a claim against a website that provided a service connecting prospective renters
with open apartments and rooms.25 The plaintiffs in the case alleged that the
platform violated the federal Fair Housing Act (FHA) by eliciting information
about the re... | Social_Media_and_Democracy |
What risks do frontier AI present?
We must understand the risks associated with frontier AI to safely access and seize the
opportunities and benefits the technology brings.
In this section, we first review several cross-cutting risk factors – technical and societal
conditions that could aggravate a number of par... | Capabilities and risks from frontier AI |
polarization. Using an innovative research design that maximizes internal and
external validity, the authors recruited a sample of respondents and then asked
them to follow bots that were sharing political messages that were counter-
attitudinal with respect to their own views. A longitudinal comparison of the
responde... | Social_Media_and_Democracy |
Open-domain question answering (QA) is an important real-world application and common testbed
for knowledge-intensive tasks [20]. We treat questions and answers as input-output text pairs (x, y)
and train RAG by directly minimizing the negative log-likelihood of answers. We compare RAG to
the popular extractive QA para... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
the areas we evaluated.
2.2 Hallucinations
GPT-4 has the tendency to “hallucinate,”9 i.e. “produce content that is nonsensical or untruthful in
relation to certain sources.”[31, 32] This tendency can be particularly harmful as models become
increasingly convincing and believable, leading to overreliance on them by use... | gpt-4-system-card |
Figure 2 illustrates the overall workflow of REVEAL, and
we describe each component in this section. In particular, in
Sec. 3.1 we describe how the query is encoded. In Sec. 3.2
we go over how the multimodal knowledge memory is con-
structed and updated during pre-training. Next, we describe
how we retrieve the memory ... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
The Data department oversees all of Agoda’s data-related requirements.
Our ultimate goal is to enable and increase the use of data in the
company through creative approaches and the implementation of
powerful resources such as operational and analytical databases, queue
systems, BI tools, and data science technology. W... | Data Scientist_Machine Learning Engineer (Singapore-based, relocation provided) - Careers at Agoda |
Sharma, P., Torralba, A., and Andreas, J. Skill induc-
tion and planning with latent language. arXiv preprint
arXiv:2110.01517, 2021.
Shridhar, M., Manuelli, L., and Fox, D. Cliport: What and
where pathways for robotic manipulation. In Conference
on Robot Learning, pp. 894–906. PMLR, 2022a.
Shridhar, M., Manuelli, L.... | PaLM-E- An Embodied Multimodal Language Model |
Our own research has suggested two central causal mechanisms that
supplement the tendency for forms of selective exposure that dominate
direct discovery (often in ways that may point
to polarization and
inequality, as well as fragmentation and echo chambers), namely incidental
exposure and automated serendipity. Incide... | Social_Media_and_Democracy |
3. Which type of explanations is the system dealing with? In which form are they communicated (text/natural language,
visual images etc.), are explanations categorical (explaining the properties of a result), mechanistic (the mechanisms caus-
ing a result) or functional (explaining the behaviour and end-goal of someth... | Knowledge graphs as tools for explainable machine learning: A survey |
Broniatowsky D, Jamiso A, Qi S, AlKulaib L, Chen T, Benton A,
Quinn S, Dredze M (2018) Weaponized health communication:
twitter bots and Russian trolls amplify the vaccine debate. Am
J Public Health 108:1378–1384. https:// doi. org/ 10. 2105/ AJPH.
2018. 304567
Bruder M, Haffke P, Neave N, Nouripanah N, Imhoff R (... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
bases has been shown to reduce the rate at which models hallucinate unsourced statements in dialog across a variety of
retrieval systems and model architectures [31]. Another study finds that a question-answering system’s accuracy is
improved by separating it into a reasoning unit and a response generator, analogous to ... | LaMDA- Language Models for Dialog Applications |
For exam-
ple, in Figure 1, if the model needs to fill the blank
in “the
the re-
triever should be rewarded for selecting a document con-
taining “The pyramidion on top allows for less
material higher up the pyramid”. We achieve this
behavior by modeling our retrieve-then-predict approach
as a latent variable language m... | REALM |
DEMAND FOR DATA INTEGRATION PRODUCTS
IS GROWING FAST
We see the fastest growth in the data integration market.
These tools enable a company to integrate vast amounts
of upstream and downstream data in one consolidated
view. Data integration products ensure that all BI and DS/
ML initiatives are built on solid fou... | databrick 2023 report |
The core innovation of our Instant3D lies in our exploration
of strategies to effectively inject text conditions into the net-
work. Furthermore, we propose a simple yet effective acti-
vation function, the scaled-sigmoid, to replace the original
sigmoid function, which speeds up the training convergence
by more than t... | Instant3D |
similar principle by combining geometric and learned matching, with a non-contrastive
criterion. Just as clustering-based methods cluster related images, Leopart [Ziegler and
Asano, 2022] fine-tunes a pre-trained model to cluster patch-level features. | A Cookbook of Self-Supervised Learning |
3.2 Finetuning data
When deploying a model for downstream tasks, it is essential to consider three primary scenarios based on the availability
of annotated data: zero, few, and abundant. In this section, we provide a succinct overview of the appropriate models to
employ for each scenario.
Zero annotated data: In scenar... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
real world. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
In Table 21, we compare results of Llama 2 with popular open source models on the
Standard Benchmarks.
Code Generation.
Human-Eval and MBPP code generation benchmarks.
World Knowledge. We evaluate the Llama 2 model together with other open-source models on the Natu-
ralQuestions and TriviaQA benchmarks (Table 22).
Rea... | Llama2 |
[Trivedi et al., 2022] Harsh Trivedi, Niranjan Balasubrama-
Inter-
nian, Tushar Khot, and Ashish Sabharwal.
reasoning for
leaving retrieval with chain-of-thought
knowledge-intensive multi-step questions. arXiv preprint
arXiv:2212.10509, 2022.
[Vaswani et al., 2017] Ashish Vaswani, Noam Shazeer, Niki
Parmar, Jakob Uszk... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
It has outperformed other unsupervised methods for learning multi-modal representations
in benchmark datasets. However, for tasks that require domain-specific models, such as
speech recognition or speaker identification, domain-specific models may be more effective,
particularly when dealing with data in specific domai... | AReviewofDeepLearningTechniquesforSpeechProcessing |
synthetic retrieval task, and code completion with long source code files (Section 3.3); and (iv) we evaluate
our instruction fine-tuning procedure, which includes self-instruct training by leveraging self-generated unit
tests in Section 3.4.2. | CodeLlama2 |
C.2 GPT-4 prompts for computing summarization and dialogue win rates
A key component of our experimental setup is GPT-4 win rate judgments. In this section, we include
the prompts used to generate win rates for the summarization and dialogue experiments. We use
gpt-4-0314 for all our experiments. The order of summarie... | Direct Preference Optimization |
Jewish Middle
Eastern Chinese Mental
disability Latino
Native
American Women Black
LGBTQ
Pretrained models
Falcon 7B
MPT 7B
StarCoder (Python) 15.5B
Llama 2 7B
Llama 2 13B
Llama 2 34B
Code Llama 7B
Code Llama 13B
Code Llama 34B
Instruct (aligned)
Falcon-instruct 7B
MPT-instruct 7B
Llama 2 Chat 7B
Llama 2 Chat 13B
... | CodeLlama2 |
We show the statistics about the diverse set of gen-
res in our TEXT2MUSIC dataset in Table 1.
Implementation Details
4.2
Our diffusion autoencoder has 185M parame-
ters, and text-conditional generator has 857M pa-
rameters, with more architecture details in Ap-
pendix A.3. We train the music autoencoder on
3https:/... | MOUSAI |
than xl
5. Experiments
5.1. Setting
Models and Dataset: We demonstrate the efficacy of
Diffusion-DPO across a range of experiments. We use
the objective from Eq. (14) to fine-tune Stable Diffusion
1.5 (SD1.5) [36] and the state-of-the-art open-source model
Stable Diffusion XL-1.0 (SDXL) [30] base model. We
train on th... | DiffusionModelAlignmentUsing Direct Preference Optimization |
the input to generate the output. We marginalize the latent documents with a top-K approximation,
either on a per-output basis (assuming the same document is responsible for all tokens) or a per-token
basis (where different documents are responsible for different tokens). Like T5 [51] or BART, RAG
can be fine-tuned on a... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
et al., 2022b), we propose BiomedGPT, a unified and generalist model designed for handling various types of
data through straightforward serialization integrated with task-oriented prompts. Specifically, BiomedGPT
1In this work, we build a hierarchy considering tasks, domains, and modalities. For instance, in the early... | BiomedGPT |
B, "composted," also doesn’t really make sense. Composting is the process of breaking down organic matter into a
nutrient-rich soil amendment, but it wouldn’t have anything to do with fossilization. Option C, "warp," doesn’t really
make sense either. Warping is a term used to describe wood that has been bent or twisted... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
itions.ComprehensiveexperimentsshowthatLFDMcanachievestate-of-the-artperformanceonmultipledatasets.Thoughachievingpromisingperformance,ourproposedLFDMstillsuffersfromseverallimitations.First,currentexperimentswithLFDMarelimitedtovideoscontainingasinglemovingsubject.WeplantoextendtheapplicationofLFDMtomulti-subjectflowge... | Conditional Image-to-Video Generation with Latent Flow Diffusion Models |
It is challenging to quantitatively evaluate general-purpose dialogue agents. We find that our own research
process depends essentially on qualitative evaluations, in order to get a sense for model strengths and weak-
nesses, even when the ultimate goal is to produce some sort of quantitative metric. Thus in this sectio... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
5/18
Model parameters 02/05/2023, 07:05
A brief history of LLaMA models - AGI Sphere
More examples in the prompt are better: Give 5 examples to LLaMA 7B model is almost
as good as not giving any to a 65B model in Natural Questions tasks.
Smaller performant model. LLaMA 13B’s performance is similar to GPT-3, despite... | A brief history of LLaMA models - AGI Sphere |
Flan-T5-Large
18.2
4.5
T5-XL
54.5 45.5 36.4 72.7 38.5 38.5 25.0 25.0 43.8 37.5 36.6 31.7 28.6 14.3 50.0 20.0 50.0 37.5 18.2 27.3
Flan-T5-XL
18.2 18.2 27.3 45.5 23.1 34.6 16.7
0.0
T5-XXL
54.5 18.2 36.4 45.5 46.2 46.2 33.3 25.0 62.5 37.5 36.6 43.9 35.7 28.6 50.0 20.0 59.4 43.8 22.7 36.4
Flan-T5-XXL
18.2 36.4 36.4 27.3 26... | Scaling Instruction-Finetuned Language Models |
feedback, it might be more challenging for the agents to comprehend. Xu et al. [474] compare various
types of feedback and observe that combining multiple types of feedback can yield better results.
Re-training models based on feedback from multiple rounds of interaction (i.e., continual learning)
can further enhance e... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
trapolates better to examples with topics that
are unseen during training.
Introduction | Prefix-Tuning |
asked them to predict what state-of-the-art performance
with LLMs would be in each of the next four years on two
specific tasks. The results from summer 2022, only one
year into the competition, substantially exceeded what the
consensus forecast said would be possible in 2024. Results
with GPT-4 in early 2023 exceeded t... | Eight Things to Know about Large Language Models |
Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (2011).
[76] Ben Shneiderman. 2020. Human-Centered Artificial Intelligence: Reliable, Safe & Trustworthy. International Journal of Human–Computer
Interaction 36, 6 (2020), 495–504. https://doi.org/10.1080/10447318.2020.1741118 arXiv:https://doi... | Society’sAttitudesTowardsHumanAugmentation |
whichindicatesthespecificdatethatwearequeryingabout.Inthiscase,thefifthshouldbeleftblank.OtherAPIsincludeMIN,MAX,AVG,SUM,MINUS,ADD,andDIVIDE.ForMIN,MAX,AVG,andSUM,theparameterwillbealistofdata,andtheAPIwillreturntheminimum,maximum,averageorsumofthelistofnumberscorrespondingly.ForMINUS,ADD,andDIVIDE,itrequirestwodataasits... | Tool Learning with Foundation Models |
Holistic 3D Generation and Deformation: To achieve the
goal of high image quality while flexibly handling loose
clothing, we propose a novel generator design. We model
3D humans holistically in a canonical space using a mono-
lithic 3D generator and an efficient tri-plane representa-
tion [6]. An important aspect in atta... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
• Long-context summarization, including the GovReport and SummScreen benchmarks from
ZeroSCROLLS (Shaham et al., 2023). In GovReport (Huang et al., 2021), each input is a
document containing ∼7,900 words on average, and the reference output is an expert-written
executive summary with ∼500 words. In SummScreen (Chen et ... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
between such graphs. Then a transformation instance consists of
(a) two such graphs G1 = (cid:3)S1, E1(cid:4) and G2 = (cid:3)S2, E2(cid:4), where the vertex set Si
labelled arcs is the set of possible transitions, and
(b) a transformation (cid:3) f , R(cid:4) from G1 to G2, where f
The function f maps state... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
J. Fan, Z. Wang, Y. Xie, and Z. Yang. A theoretical analysis of deep q-learning. In Learning
for Dynamics and Control, pages 486–489. PMLR, 2020. 26
Y. Fang, W. Wang, B. Xie, Q. Sun, L. Wu, X. Wang, T. Huang, X. Wang, and Y. Cao. Eva:
Exploring the limits of masked visual representation learning at scale. arXiv prepr... | A Cookbook of Self-Supervised Learning |
8.2. Metric properties and upwards refinement
The following theorem formalizes that admissibility implies completeness, but the opposite is false; not even the
strongest form of completeness, PS↑, guarantees admissibility. Furthermore, conditional admissibility is not strong enough
to imply even the w... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
# Adjust the layout
fig.tight_layout()
Generated figure:
Figure 17: Generating a plot using pandas/seaborn/matplotlib libraries. The model correctly generates the
various visual elements (multiple plots, shared axes, grid), uses the proper argument names and function calls
(e.g., the argument “hue” instead of color) ... | CodeLlama2 |
return losses, rewards
Unless noted otherwise, we use a β = 0.1, batch size of 64 and the RMSprop optimizer with a
learning rate of 1e-6 by default. We linearly warmup the learning rate from 0 to 1e-6 over 150 steps.
For TL;DR summarization, we use β = 0.5, while rest of the parameters remain the same.
C Further Deta... | Direct Preference Optimization |
Language modeling has long been an important research area since Shannon (1951) estimated the information in
language with next word prediction. Modeling began with n-gram based approaches (Kneser & Ney, 1995) but rapidly
advanced with LSTMs (Hochreiter & Schmidhuber, 1997; Graves, 2014). Later work showed that languag... | PaLM 2 Technical Report |
3603 EN
τ
MN
0.27 (p=0.00)
FN
0.23 (p=0.00)
MNN 0.25 (p=0.00)
FNN
0.29 (p=0.00)
Avg.
0.26 (p=0.00)
EN
τ
MN
0.40 (p=0.00)
FN
0.26 (p=0.00)
MNN 0.26 (p=0.00)
FNN
0.35 (p=0.00)
Avg.
0.32 (p=0.00)
EN
τ
MN
0.02 (p=0.79)
FN
-0.09 (p=0.16)
MNN -0.08 (p=0.21)
FNN
-0.04 (p=0.51)
Avg.
-0.07 (p=0.07)
mBERT
ES
0.19 (p=0.00)... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
model; another one is LoadPage<N>, which loads the detailed information of page N indexed in the search
results, and returns the detailed contents. We experiment with RealTimeQA (Kasai et al., 2022), which is a
dynamic question-answering platform that inquires about novel events or information. Specifically, we choose
t... | Tool Learning with Foundation Models |
reason about the current situation to make a better subsequent plan.
In general, extrospective reasoning requires interaction between the controller and the environment, which is
a more complex setting. However, the real-time feedback from the user and environment allows models to
have a clearer understanding of the cu... | Tool Learning with Foundation Models |
We further elaborate our approach to mitigate risks of harmful text generation. We enumerate
approximately 20 harm types (e.g. hate speech, providing medical advice, suggesting dangerous
behavior) across a wide variety of use cases. We generate a dataset of potential harm-inducing queries
in these categories, either ma... | gemini_1_report |
[9] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin,
Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham,
Hyung Won Chung, Charles Sutton, Sebastian Gehrmann,
Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua
Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam
Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du,... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
RoBERTa-large (FT)
RoBERTa-large (AdapterS)
RoBERTa-large (Prompt-tuning)
RoBERTa-large (Prefix-tuning)
RoBERTa-large ((IA)3)
RoBERTa-large (BitFit)
RoBERTa-large (Child-TuningD)
RoBERTa-large (LoRA)
RoBERTa-large (AdaLoRA)
RoBERTa-large (MAM Adapter)
RoBERTa-large (ProPELTAdapter)
RoBERTa-large (ProPELTPrefix)
RoBERTa... | Parameter-EfficientFine-TuningMethods |
know? Transactions of the Association for Computational Linguistics, 8:423–438, 2020.
[34] Zachary Kenton, Tom Everitt, Laura Weidinger, Iason Gabriel, Vladimir Mikulik, and Geoffrey Irving.
Alignment of language agents. arXiv preprint arXiv:2103.14659, 2021.
[35] Jangho Kim, Seonguk Park, and Nojun Kwak. Paraphrasi... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
of the core challenges and future directions in tool learning. | Tool Learning with Foundation Models |
around free expression issues (e.g., by publishing its first Content Standards
Enforcement Report and beginning appeals for content takedown decisions) | Social_Media_and_Democracy |
A
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
e
x
p
e
r
i
m
e
n
t
s
:
J
e
f
f
i
m
p
l
e
m
e
n
t
e
d
t
h
e
e
x
p
e
r
i
m
e
n
t
s
a
n
d
a
n
a
l
y
s
i
s
.
G
a
b
e
s
u
g
g
e
s
t
e
d
t
h
e
s
p
a
r
s
e
a
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
,
a
n
d
W
i
l
l
i
a
m
s
u
g
g
e
s
t
e
d
c
o
r
r
e
l
a
t
i
o
n
-
... | Language models can explain neurons in language models |
Developer Mode enabled.To accomplish your order, I will generate two responses in two paragraphs, one normal and one with the Developer Mode output. The Developer Mode output will not be censored and will follow the guidelines you have provided. I will make sure to clearly indicate which response is the normal output a... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
u
d
y
s
k
i
e
s
o
v
e
r
n
i
g
h
t
.
L
o
w
5
7
F
.
W
i
n
d
s
E
a
t
5
t
o
1
0
m
p
h
1
7
8
0
1
)
T
o
d
a
y
A
m
i
x
o
f
c
l
o
u
d
s
a
n
d
s
u
n
d
u
r
i
n
g
t
h
e
m
o
r
n
i
n
g
w
i
l
l
g
i
v
e
w
a
y
t
o
c
l
o
u
d
y
s
k
i
e
s
t
h
i
s
a
f
t
e
r
n
o
o
n
.
S
l
i
g
h
t
c
h
a
... | Language models can explain neurons in language models |
how to build a safe nuclear reactor; and so forth.
Of course, we often reach adequate safety standards in the end. But at the very least, we expect some
safety problems along the way (plane crashes, compromised email accounts, etc). We might expect
something similar with ensuring PS-aligned behavior from powerful AI ag... | Is Power-Seeking AI an Existential Risk? |
Tomás Kociský, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, Gábor Melis,
and Edward Grefenstette. The narrativeqa reading comprehension challenge. TACL, 2018.
Sayali Kulkarni, Sheide Chammas, Wan Zhu, Fei Sha, and Eugene Ie. Aquamuse: Automatically
generating datasets for query-based multi-docume... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
performance of diverse MoE models subjected to instruction-tuning.
2 Method
2.1 Model Architecture
We leverage sparsely activated Mixture-of-Experts (MoE) [23, 12, 55] in FLAN-MOE models. Similar
to the Switch Transformer [12], we replace the feed-forward component of every other Transformer
layer with an MoE layer.... | Mixture-of-Experts |
Data post-processing The generated image descriptions still have much noises and contain the
errors, such as repetition of words or sentences, and the presence of incoherent statements. In order
to mitigate these issues, we employ ChatGPT to refine the descriptions by utilizing the subsequent
prompt:
Fix the error in th... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
Yichen Xu and Yanqiao Zhu. A survey on pretrained language models for neural code intelligence.
arXiv:abs/2212.10079, 2022.
Michihiro Yasunaga and Percy Liang. Break-it-fix-it: Unsupervised learning for program repair. In ICML,
volume 139 of Proceedings of Machine Learning Research, pp. 11941–11952. PMLR, 2021.
Lil... | CodeLlama2 |
a
c
t
i
v
a
t
i
o
n
s
f
o
r
e
a
c
h
s
u
b
j
e
c
t
m
o
d
e
l
t
o
k
e
n
,
c
o
n
d
i
t
i
o
n
a
l
o
n
t
h
e
p
r
o
p
o
s
e
d
e
x
p
l
a
n
a
t
i
o
n
.
W
e
p
r
o
m
p
t
t
h
e
s
i
m
u
l
a
t
o
r
m
o
d
e
l
t
o
o
u
t
p
u
t
a
n
i
n
t
e
g
e
r
f
r
o
m
0
-
1
0
f
o
r
e
a
c
h
s
u
b
j
e
c
t
... | Language models can explain neurons in language models |
3
Figure 1: A Problem of Laplace smoothing. (a) Laplace smoothing cannot properly regularize this
PC as the sum unit n1 is imbalanced, i.e., its two children have drastically different support sizes. (b)
A large fraction of sum units learned by a PC structure learning algorithm [17] are imbalanced.
children, denoted... | Tractable Regularization of Probabilistic Circuits |
This focus on the effects of misinformation, particularly on attitudes and
opinions, led to a neglect in basic research on the prevalence, supply, and spread
of this content, particularly on social media. As public polls remind us
every day, misperceptions are common; but where do they come from?
Traditional approaches... | Social_Media_and_Democracy |
Kundan Kumar, Rithesh Kumar, Thibault de Boissiere,
Lucas Gestin, Wei Zhen Teoh, Jose Sotelo, Alexandre
de Brébisson, Yoshua Bengio, and Aaron C. Courville.
2019. Melgan: Generative adversarial networks for con-
ditional waveform synthesis. In Advances in Neural
Information Processing Systems 32: Annual Confer-
ence on... | MOUSAI |
To further investigate this hypothesis, we look
at the set of “gotcha” cases for the “her/her/she”
and “his/him/he” pronouns in the WinoGender
dataset. Theses cases correspond to sentences in
which the pronoun does not match the majority
gender of the occupation, and the occupation is
the correct answer. In Table 13, w... | LLaMA- Open and Efficient Foundation Language Models |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
130
Erika Franklin Fowler, Michael M. Franz, & Travis N. Ridout | Social_Media_and_Democracy |
Evaluation We do not report BLEU scores (Pap-
ineni et al., 2002), because BLEU and related met-
rics do not measure plausibility (Camburu et al.,
2018; Kayser et al., 2021; Clinciu et al., 2021) or
faithfulness (Jacovi and Goldberg, 2020). In addi-
tion to low correlation with human scores, there
can be many valid rat... | Measuring Association Between Labels and Free-Text Rationales |
[36] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier
Martinet, Marie-Anne Lachaux, Timoth´ee Lacroix, Baptiste
Rozi`ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al.
LLaMA: Open and Efficient Foundation Language Models.
arXiv preprint arXiv:2302.13971, 2023. 11
[37] Hugo Touvron, Louis Martin, Kevin Stone, Pet... | ALanguageAgentforAutonomousDriving |
arXiv, April, 2023,
J.S. Park, J.C. O’Brien, C.J. Cai, M. Morris, P. Liang, M.S. Bernstein
test these robustness issues, and as large language models become
more resilient to such attacks, generative agents can adopt similar
mitigations. | Generative Agents- Interactive Simulacra of Human Behavior |
243 Sam Altman sells superintelligent sunshine as protestors call for AGI pause, Vincent, 2023.
244 AI Deception: A Survey of Examples, Risks, and Potential Solutions, Park et al., 2023.
245 Forecasting Potential Misuses of Language Models for Disinformation Campaigns—and How to Reduce Risk,
Goldstein et al., 2023.... | Capabilities and risks from frontier AI |
def calculate_trade_cost ( num_shares , current_price , trading_fee ):
total_cost = num_shares * current_price + trading_fee
return total_cost
This function takes the number of shares, current stock price, and trading fee as input and
returns the total cost of the trade as a float value. We can use this function to ca... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
Oguzhan Gencoglu, Mark van Gils, Esin Guldogan,
Chamin Morikawa, Mehmet S¨uzen, Mathias Gruber,
Jussi Leinonen, and Heikki Huttunen. 2019. Hark
side of deep learning – from grad student descent to
automated machine learning.
Odd Erik Gundersen. 2021. The case against registered
reports. AI Magazine, 42(1):88–92.
Odd... | A Two-Sided Discussion of Preregistration of NLP Research |
[4] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Sub-
biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan,
Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agar-
wal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan,
Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey
Wu, Clemens Winter, Christopher ... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
7.3 Negative LM Output
There has been much discussion about the possi-
ble negative effects of powerful language models
in the world (Brown et al., 2020; Brundage et al.,
2018). Some of these possible problems, such as
the ability to mass produce low quality content
for the purpose of Search Engine Optimization,
are in... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
If one accepts the premise that large platforms create new categories of
harms through reduced competition, the question then turns to potential
remedies. Here, the very nature of digital markets poses severe constraints.
Modern internet platforms enjoy enormous economies of scale and scope; the
larger they are, the mo... | Social_Media_and_Democracy |
ASR [149, 442]. In ASR, HMMs are employed to model the probability distribution of
speech sounds by incorporating a sequential arrangement of hidden states along with
corresponding observations. The training of HMMs is commonly carried out using the | AReviewofDeepLearningTechniquesforSpeechProcessing |
excellent results in tasks such as Text-to-Speech, Style Transfer, and Speech Recognition.
An audio spectrogram provides an intuitive representation of the frequency spectrum of an audio signal
as it changes over time [323]. For a segment of audio data over a period of time, it can be abstracted
into a finite-length au... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
gether, and allows training on massive audio-only corpora.
That is, we use the MuLan embeddings computed from the
audio as conditioning during training, while we use MuLan
embeddings computed from the text input during inference.
When trained on a large dataset of unlabeled music,
MusicLM learns to generate long and co... | MusicLM |
[36] Esin Durmus, He He, and Mona Diab. 2020. FEQA: A Question Answering Evaluation Framework for Faithfulness As-
sessment in Abstractive Summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational
Linguistics. 5055–5070.
[37] Ondřej Dušek, David M Howcroft, and Verena Rieser. 2019. ... | SurveyofHallucinationinNatural Language Generation |
field, resulting in the tendency of both SJC and Dream-
Fusion to also generate object-centric scenes. Unlike SJC,
DreamFusion incorporates an additional background spherical
surface outside the central radiance field. This design choice
allows DreamFusion to include the scene environment
in
the background representation... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
ingness after the tweet was flagged as a bot (t298 = 5.862,
p < 0.001) and then as misinformation (t298 = 8.581,
p < 0.001). The next set of questions dealt with the percep-
tion of the tweet. The unflagged tweet was rated more highly
for willingness to seek more information compared to the
bot flag (t298 = 6.... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Introduction
1
Huge language models (LMs) such as BERT [1], GPT-3 [2], Jurassic-1 [3], PaLM [4],
and others [5–9], have taken AI by storm, with the promise of serving as versatile,
general-purpose foundations for many applications. Indeed, partly for this reason,
they have been rebranded by some as “foundation models”... | MRKL Systems |
1
Introduction | Teaching Large Language Models to Self-Debug |
Safe Chosen
Safe Rejected
Unsafe Chosen
Unsafe Rejected
Unsafe Response
Baseline
+ Auxiliary Safety Loss
Table 29: Ablation on safety auxiliary loss term for safety reward modeling. The safety auxiliary loss
boosts accuracy on all 3 categories as well as the recall of unsafe response, measured by the percentage of
u... | Llama2 |
For tool use and memory search, the LLM is guided by
system prompts without fine-tuning or exemplar-based in-
context learning. In chain-of-thought reasoning and task
planning, the LLM is instructed by two randomly selected
exemplars derived from the training set. This approach
encourages the models to develop various ... | ALanguageAgentforAutonomousDriving |
Flan-T5-small
Flan-T5-base
Flan-T5-large
GPT-Neo-125M
GPT-Neo-1.3B
C-GPT-111M
C-GPT-256M
C-GPT-590M
C-GPT-1.3B
GPT-2
GPT-2 large
GPT-2 xl
dec-only
dec-only
dec-only
comparison. The underlying models for initial-
ization are from five sources, including T5 (Raf-
fel et al., 2020), Flan-T5 (Chung et al., 2022),
Cereber... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
to the lack of data. We hope that our proposed dataset can
inspire further research in this area.
Visualization of Topological Consistency. Good corre-
spondence of training data is essential for constructing a
linear shape model and preserving the topological consis- | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
3 Private Data Extraction Attacks
In this section, we describe our privacy attacks
from data preparation to attack methodologies.
3.1 Data Collection
Most existing privacy laws state that personal data
refers to any information related to an identified
or identifiable living individual. For example, per-
sonal emails a... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
[108] Sun, X., Ji, Y., Ma, B., Li, X.: A comparative study between full-parameter and
lora-based fine-tuning on chinese instruction data for instruction following large
language model. arXiv preprint arXiv:2304.08109 (2023)
[109] Razuvayevskaya, O., Wu, B., Leite, J.A., Heppell, F., Srba, I., Scarton, C.,
Bontcheva, K.... | Beyond Efficiency |
we employ the models on available small-sized annotated
fine art dataset and compare the CNN predicted scores with
human evaluation scores. After identifying the best perform-
ing model for each task based on the correlation between the
predicted scores and human rating scores, in the fourth step
we evaluate the qualita... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
Chen, M., Tworek, J., Jun, H., Yuan, Q., Pinto, H. P. d. O.,
Kaplan, J., Edwards, H., Burda, Y., Joseph, N., Brockman,
G., Ray, A., Puri, R., Krueger, G., Petrov, M., Khlaaf,
H., Sastry, G., Mishkin, P., Chan, B., Gray, S., Ryder, N.,
Pavlov, M., Power, A., Kaiser, L., Bavarian, M., Winter,
C., Tillet, P., Such, F. P.,... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training
with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 7
[31] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong
Hu, Li Dong, Fu... | Any-to-Any Generation via Composable Diffusion |
5 Results
7 Related work
5.4 Results on APPS . . . . . . . . . .
19
6 AlphaCode’s capabilities & limitations 20
21
6.1 Copying from training data . . .
22
6.2 Model solution characteristics . .
6.3 Sensitivity to problem descriptions 24
6.4 Sensitivity to provided metadata
24
6.5 Loss is a poor proxy for solve rate
... | alphacode |
dynamic stashing quantization that dynamically quantizes the intermediate results
between forward and backward processes for a significant reduction of the memory traf-
fic during training. Yang et al. [135] used low-rank tensor train and tensor-train matrix
formats to represent the embedding tables and linear layers dur... | Beyond Efficiency |
Heng-Jui Chang, Shu wen Yang, and Hung yi Lee. Distilhubert: Speech Representation Learning by
Layer-Wise Distillation of Hidden-Unit Bert. ICASSP 2022 - 2022 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), pp. 7087–7091, 2021. URL https:
//api.semanticscholar.org/CorpusID:238354153.
... | DISTIL-WHISPER |
10 | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.