text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
2022) just as much as our baseline.
In general, our kernels target generative inference in the low batch-size setting (for simplicity, we
consider only batchsize 1) where the underlying (close to) matrix-vector products are memory-
bound. For non-generative and large-batch applications, operations may be compute- rathe... | GPTQ |
4 Experiments
4.1 Datasets | DOCLLM |
77.7%
3-shot
—
—
82.0
Variable
shots
86.8%
10-shot
72.7
1-shot
88.0%
0-shot
—
—
—
—
—
—
Table 2 | Gemini performance on text benchmarks with external comparisons and PaLM 2-L.
∗ The model produces a chain of thought with k = 8 or 32 samples, if there is a consensus above a threshold (chosen based on the va... | gemini_1_report |
Bode, L., Lassen, D. S., Kim, Y. M. et al. (2016). Coherent campaigns? Campaign
broadcast and social messaging. Online Information Review, 40(5), 580–594.
Borah, P. (2016). Political Facebook use: Campaign strategies used in 2008 and 2012
presidential elections. Journal of Information Technology & Politics, 13(4),
32... | Social_Media_and_Democracy |
1. Does the literature review discuss about authenticity of his problem statement?
2. Does the literature review significantly support the severity of his problem statement?
3. Does the researcher agree or disagree with existing knowledge, and why?
4. Is his/her final judgment or conclusion is sound, logical and... | How to Write Your PhD Proposal- A Step-By-Step Guide |
Both professionals and non-professionals can access LLMs online [28]. In this context, our analysis of typical usage
scenarios and differences between novice and expert users sheds light on contexts that need to be prioritized for
supporting effective use across domains and experience levels.
2.3 Interaction with LLMs... | Adoptionand AppropriationofLLMs |
23
to become truly powerful. Our experiments shed
some initial light on the internal mechanisms of
LLMs and start to unravel how these models are ca-
pable of performing numerous tasks. For example,
despite their zero-shot capabilities, in the absence
of explicit examples (or in the case where the in-
structions were... | AreEmergentAbilitiesinLarge Language Models just In-Context |
from improving the quality of our generated data by
using human annotators or training a reward model
to select better generations, similar to the algorithm
used in Ouyang et al. (2022).
5.5 Example Predictions from GPT3SELF-INST
We present a selection of user-oriented tasks, the
corresponding GPT3SELF-INST-produced re... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
:
I
t
o
p
e
r
a
t
e
s
o
n
t
h
e
a
s
s
u
m
p
t
i
o
n
t
h
a
t
i
n
h
i
g
h
d
i
m
e
n
s
i
o
n
a
l
s
p
a
c
e
,
d
i
s
t
a
n
c
e
s
b
e
t
w
e
e
n
n
o
d
e
s
f
o
l
l
o
w
a
G
a
u
s
s
i
a
n
d
i
s
t
r
i
b
u
t
i
o
n
a
n
d
t
h
u
s
t
h
e
r
e
s
h
o
u
l
d
e
x
i
s
t
c
l
u
s
t
e
r
i
n
g
o
f
d
... | LLM Powered Autonomous Agents _ Lil'Log |
anonymization given the richness of those datasets and the outside chance that
researchers might theoretically be able to reidentify people if they were
committed to combining multiple datasets from other sources. | Social_Media_and_Democracy |
every z ∈ Z and construct an efficient search index over
these embeddings. However, this data structure will no
longer be consistent with p(z | x) if the parameters θ of
Embeddoc are later updated. Hence, the search index goes
“stale” after every gradient update on θ.
Our solution is to “refresh” the index by asynchrono... | REALM |
12
Approach
Time Complexity Memory Complexity
O(T d log T )
O(T log T + T d)
Transformer [11]
Reformer [45]
Linear Transformer [46]
Efficient Attention [47]
AFT [48]
Memory Efficient Attention [49]
KDEformer [51]
MEGA [52]
RWKV [53]
O(T 2d)
O(T d2)
O(T 2d)
O(T d)
O(T 2d)
O(mT d)
O(cT d)
O(T d)
O(T 2 + T d)
... | Beyond Efficiency |
Amendment of Section 230
271
discriminatory content posted on that portion of the site. Extending such a rule
to the political context would mean that platforms might, for instance, be
granted immunity for activities occurring on its freeform posting features but
not
for algorithmic feeds where the platform plays a m... | Social_Media_and_Democracy |
When viewed this way, it becomes clear that, despite their value, current LMs
have inherent limitations. While versatile and impressive, the output of even huge
LMs is in many cases wrong, and often ridiculously so [13]. Here is a sample output
of GPT-3 on some simple queries. (To be clear, this is not a critique of GP... | MRKL Systems |
4 . 3
P I I D E T E C T I O N M O D E L
We fine-tuned StarEncoder on the annotated PII dataset for the Named Entity Recognition (NER) task.
We added a linear layer as a token classification head on top of the model, with 6 target classes: names,
emails, keys, passwords, IP addresses, and usernames. We excluded IDs du... | StarCoder_paper (1) |
Furthermore, we show more examples of 3D scenes gen-
erated by our Text2NeRF in Fig. 7. It is worth noting that
our method can not only generate diverse results from the
same text prompt (Fig. 7(g)&(j), (h)&(k), and (i)&(l)), but also
support to generate 3D scenes in some artistic styles (Fig. 8).
Please refer to the s... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
3 Creating AI Art
3.1 Technological Milestones | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
vocabulary to represent the generation of all modalities as a
language modeling problem. This produces a total vocabu-
lary size of approximately 300,000.
3.3. Super-Resolution
Generating high-resolution (HR) videos with an autoregres-
sive transformer incurs heavy computational cost due to the
increase in sequence len... | VideoPoet |
1018
R
N
(cid:11)
Transformation
S
(cid:9)
substitute with similar info.
substitute with dissimilar info. (cid:9)
≡
paraphrasing
(cid:2) (cid:2) (cid:2)
negation
(cid:9) (cid:9) (cid:9)
transform to specific
(cid:11) (cid:11) (cid:11)
transform to general
Table 2: NatOp assignment based on transfor-
mations and ve... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
sum = 0
for i in range(1,((n + 1))):
for j in range(i,((n + 1))):
sum = (sum + (i * j))
return sum
[/python]
[c++]
<insert C++ program here>
[/c++]
[python]
<insert original Python translation here>
[/python]
The above Python translation does not do the same thing as the C++ code.
Correct the Python translation.
[p... | Teaching Large Language Models to Self-Debug |
Language models can explain neurons in language models
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
32/32 | Language models can explain neurons in language models |
Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong
Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to
follow instructions with human feedback. arXiv preprint arXiv:2203.02155.
Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021. Are ... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
1While the paper discusses the Pile at length, it does not ex-
plicitly state that Jurassic-1 was trained on the Pile. We originally
discovered this fact by executing data extraction attacks on the
API, and confirmed with private communication with the authors.
Pythia: A Suite for Analyzing Large Language Models
GPT-... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
ViViT [1] with additional architectural changes for video generation, which can compress the videos
in temporal and spatial dimensions, while staying auto-regressive in time. This capability allows for
generating videos of arbitrary length auto-regressively. | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
[100] K. Shu, L. Cui, S. Wang, D. Lee, and H. Liu, ‘‘DEFEND: Explainable
fake news detection,’’ in Proc. 25th ACM SIGKDD Int. Conf. Knowl.
Discovery Data Mining, Jul. 2019, pp. 395–405.
[101] M. Potthast, J. Kiesel, K. Reinartz, J. Bevendorff, and B. Stein,
‘‘A stylometric inquiry into hyperpartisan and fake news,’’ 2... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
The duck did not like the sm elly pond.
The new pond was not smelly.
Lucy loved to play outside under the big sky.
He suggested, ”Let’s play a game to forget the scary wind.”
One day, a girl named Amy wanted to have a fun day with
her friends.
Once upon a time, there was a modest girl named Sue.
On the mountain, ... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
Figure 13. The three proposed designs of motion decomposition. We choose design 3 (c) as a result of best quality of novel view synthesis,
shown in Fig. 14.
erwise always appear as the second argument to each of
T, Tskel, TNR.)
(1) Both Tskel and TNR conditioned on an observed point
position x, illustrated in Fig. 1... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
0-shot 1-shot 5-shot 64-shot
Gopher
280B 43.5
Chinchilla 70B 55.4
7B 50.0
13B 56.6
33B 65.1
65B 68.2
LLaMA
-
-
53.4
60.5
67.9
71.6
57.0
64.1
56.3
63.1
69.9
72.6
57.2
64.6
57.6
64.0
70.4
73.0
Table 5: TriviaQA. Zero-shot and few-shot exact
match performance on the filtered dev set.
3.3 Reading Comprehension
We ev... | LLaMA- Open and Efficient Foundation Language Models |
ACM Comput. Surv., Vol. 1, No. 1, Article . Publication date: February 2022.
26
Ziwei Ji, et al.
Early works on GQA mostly tried to improve the faithfulness of the answer by investigating
reliable external knowledge sources or incorporating multiple information sources. Yin et al. [220]
propose Neural Generative Qu... | SurveyofHallucinationinNatural Language Generation |
preprint arXiv:2006.09011, 2020.
[57] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal
Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. WaveNet: A generative model for raw audio.
arXiv preprint arXiv:1609.03499, 2016.
[58] Aaron van den Oord, Nal Kalchbrenner, and Kor... | Denoising Diffusion Probabilistic Models |
parameters with specific entities, it only needs
to access a fraction of its parameters at infer-
ence time, and we show that the correct identi-
fication and representation of entities is essen-
tial to EAE’s performance. | Entities as Experts- Sparse Memory Access with Entity Supervision |
Brian A. Nosek, Charles R. Ebersole, Alexander C.
DeHaven, and David T. Mellor. 2018. The prereg-
istration revolution. Proceedings of the National
Academy of Sciences, 115(11):2600–2606.
Mark Parascandola. 2010. Epistemic risk: empirical
science and the fear of being wrong. Law, Probabil-
ity and Risk, 9(3-4):201–214... | A Two-Sided Discussion of Preregistration of NLP Research |
The goal of this book is to synthesize the existing research on social media
and democracy. We present reviews of the literature on disinformation,
polarization, echo chambers, hate speech, bots, political advertising, and new
media. In addition, we canvass the literature on reform proposals to address the
widely perce... | Social_Media_and_Democracy |
will be beneficial. Finally, new benchmarks, e.g. our emer-
gent zero-shot task to measure emergent abilities of multi-
modal models, would help create exciting new applications.
Our model is a research prototype and cannot be readily
used for real world applications ( Appendix F).
Acknowledgements: Authors would like ... | IMAGEBIND- One Embedding Space To Bind Them A |
Ours
(diffuse albedo) (specular albedo) (normals)
PSNR
18.30
22.47
PSNR
19.77
27.17
PSNR
27.26
26.69 | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
cosine functions to represent the differences between tokens. Sandwich leverages the periodic nature of cosine functions to | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
learning has seen many related methodological dis-
cussions (Gencoglu et al., 2019; Lipton and Stein-
hardt, 2018; Gundersen et al., 2022), there has,
to the best of our knowledge, been no published
discussions of preregistration practice in this field,
with the exception of Gundersen (2021).89
In our discussion below,... | A Two-Sided Discussion of Preregistration of NLP Research |
S. Reed, K. Zolna, E. Parisotto, S. G. Colmenarejo, A. Novikov, G. Barth-maron, M. Giménez,
Y. Sulsky, J. Kay, J. T. Springenberg, T. Eccles, J. Bruce, A. Razavi, A. Edwards,
N. Heess, Y. Chen, R. Hadsell, O. Vinyals, M. Bordbar, and N. de Freitas. A Generalist
Agent. Transactions on Machine Learning Research, Nov. 202... | A Cookbook of Self-Supervised Learning |
+ | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Retrieval-augmented generation with a huge frozen LM (Section 3).
In the open-book variant
of the open-domain question-answering setting, the answer generator typically attends to 100+
retrieved documents, and is therefore called a reader. Current readers are fine tuned for this long-
context functionality. Because it i... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
1 https://minecraft.fandom.com/wiki/Minecraft_Wiki
2 https://minecraft.fandom.com/wiki/Tutorials/Organization#Categories
8
JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models
Table 2: Results of JARVIS-1 and baselines on Minecraft. The detailed task instructions, settings and res... | JARVIS-1 |
[50] Pablo Palafox, Aljaˇz Boˇziˇc, Justus Thies, Matthias Nießner,
and Angela Dai. NPMs: Neural parametric models for 3D
deformable shapes. In Proceedings of the IEEE/CVF Inter-
national Conference on Computer Vision, 2021. 1
[51] Jeong Joon Park, Peter Florence, Julian Straub, Richard
Newcombe, and Steven Lovegrove.... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
joint speech activities of all speakers for each segment [145]. Following the trend, several
other works propose enhanced architectures based on self-attention [324, 630] | AReviewofDeepLearningTechniquesforSpeechProcessing |
By relying on pretrained and frozen MuLan, we need audio-
only data for training the other components of MusicLM.
We train SoundStream and w2v-BERT on the Free Music
Archive (FMA) dataset (Defferrard et al., 2017), whereas
the tokenizers and the autoregressive models for the seman-
tic and acoustic modeling stages are ... | MusicLM |
Automated Signature Generation for Network Intrusion Detection Systems
(NIDS)
Supervisor:Dr Fabio Pierazzi
A Network Intrusion Detection System (NIDS) is a probe that passively monitors network traffic
and triggers a “security alert” whenever a signature matching a particular pattern is found. However,
signatures... | informatics-phd-projects-2022-23 |
agnostic. Not only does DALL·E 2 also integrate some specific cultural memories (e.g. being
able to reproduce images in the style of some well-known artists), but it also expands cultural
dependency from the training data to the choice of filtering the training data. The way DALL·E
2 pre-training mitigations wer... | The Myth of Culturally Agnostic AI Models |
4
Fig. 2. Overview of our Text2NeRF. Given an input text prompt, we infer an initial view I0 and estimate its depth D0 via a pre-trained diffusion model
and a depth estimation model. Then we use the depth image-based rendering (DIBR) to warp the initial view and its depth map to various views to build
the support set ... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
Zoph, B. and Le, Q. V. Neural architecture search with
reinforcement learning. In ICLR, 2017.
McCann, B., Bradbury, J., Xiong, C., and Socher, R.
Learned in translation: Contextualized word vectors. In
Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fer-
gus, R., Vishwanathan, S., and Garnett, R. (eds.), NIPS.
20... | Parameter-Efficient Transfer Learning for NLP |
Navigating Knowledge Graphs | Tool Learning with Foundation Models |
parameters of a language model? arXiv preprint arXiv:2002.08910, 2020.
Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: BM25 and beyond.
Found. Trends Inf. Retr., 3(4):333–389, April 2009. ISSN 1554-0669. doi: 10.1561/1500000019.
URL https://doi.org/10.1561/1500000019.
Victor Sanh, Albert W... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
random and hypervariable sequences. It is just that using the 20 amino acid code is not a suitable
approach to explain the phenomenon. Therefore, herein, we will calculate more than 430 different
physicochemical properties to represent each residue of all antibodies, in an effort to identify what is
the right dictio... | informatics-phd-projects-2022-23 |
[228] Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing
Dialogue Agents: I have a dog, do you have pets too?. In Proceedings of the 56th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers). 2204–2213.
[229] Tianyi Zhang, Vars... | SurveyofHallucinationinNatural Language Generation |
of cooperative actions by other agents or humans. In comparison to pure agent cooperation, we desire
human involvement for two main reasons: first, to ensure interpretability, as interactions between
pure agents could generate incomprehensible language [495]; second, to ensure controllability, as the
pursuit of agents ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
1026
Hong Kong, China. Association for Computa-
tional Linguistics. https://doi.org/10
.18653/v1/K19-1046
Andreas Hanselowski, Hao Zhang, Zile Li, Daniil
Sorokin, Benjamin Schiller, Claudia Schulz,
and Iryna Gurevych. 2018. UKP-athene: Multi-
sentence textual entailment for claim verifi-
cation. In Proceedings of th... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
if any, amendments
should be made to section 230 of | Social_Media_and_Democracy |
MuLan Cycle Consistency (MCC). As a joint music-
text embedding model, MuLan can be used to quantify the
similarity between music-text pairs. We compute the MuLan
embeddings from the text descriptions in MusicCaps as
well as the generated music based on them, and define the
MCC metric as the average cosine similarity be... | MusicLM |
Note, though, that compressing an argument into very few premises (or just directly forecasting the
conclusion) risks hiding conjunctiveness, too.174
As an initial step in attempting to combat (II), and possibly (I), I’ve added a short appendix where I
reformulate the argument using fewer premises; and to combat some o... | Is Power-Seeking AI an Existential Risk? |
(prior best of 41.5 from Karpukhin et al. (2020)). We find significant improvements on adversari-
ally constructed datasets (ANLI R3 and WinoGrande XL). ANLI R3 (Nie et al., 2019) improves the
state-of-the-art to 74.7 (prior best of 53.4).
We note some weaknesses in our model. ST-MoE-32B has lackluster performance on the... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
{peiyuan_zhang, tianduo_wang, luwei}@sutd.edu.sg
guangtao_zeng@mymail.sutd.edu.sg
Abstract
We present TinyLlama, a compact 1.1B language model pretrained on around 1
trillion tokens for approximately 3 epochs. Building on the architecture and tok-
enizer of Llama 2 (Touvron et al., 2023b), TinyLlama leverages variou... | TinyLlama |
the prompts, selecting the response that is safest according to a set of guidelines. We then use the human
preference data to train a safety reward model (see Section 3.2.2), and also reuse the adversarial prompts to
sample from the model during the RLHF stage. | Llama2 |
Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, H. Francis Song,
John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan,
Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks,
Maribeth Rauh, Po-Sen Huang, Amelia Gla... | gemini_1_report |
OpenAI. GPT-4 technical report. arXiv preprint arXiv:2009.03300, 2023a. (cited on pp. 1 and 4)
OpenAI. GPT-4 system card. https://cdn.openai.com/papers/gpt-4-system-card.pdf, 2023b.
(cited on p. 33)
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for auto-
In Proceedings of the 40th Annua... | StarCoder_paper (1) |
3.2 Curating a high-quality alignment dataset for vision-language domain.
To achieve greater naturalness in the generated language and enhance the model’s usability, a second-
stage alignment process is essential. While in the realm of NLP, instruction fine-tuning datasets
[30] and conversations [1] are easily accessib... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
• Understanding the Interplay among Different Tools. The multi-step multi-tool scenario typically
involves a complex task, which demands a higher level of intent understanding and reasoning capabil-
ity. To effectively utilize multiple tools under this scenario, models should not only understand tools’
individual funct... | Tool Learning with Foundation Models |
A Review of Deep Learning Techniques for Speech Processing
73 | AReviewofDeepLearningTechniquesforSpeechProcessing |
r
u
n
c
a
t
i
n
g
b
a
s
e
d
o
n
m
a
g
n
i
t
u
d
e
o
f
c
o
e
f
f
i
c
i
e
n
t
,
w
h
i
c
h
r
e
s
u
l
t
e
d
i
n
e
v
e
n
p
o
o
r
e
r
s
c
o
r
e
s
.
[
↩
]
| Language models can explain neurons in language models |
[480] Pascal Scalart et al. 1996. Speech enhancement based on a priori signal to noise estimation. In 1996 IEEE International
Conference on Acoustics, Speech, and Signal Processing Conference Proceedings, Vol. 2. IEEE, 629–632.
A Review of Deep Learning Techniques for Speech Processing
103
[481] Carolina Scarton, ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
similar to those contained in HumanEval and MBPP. | CodeLlama2 |
3
2
0
2
r
p
A
7
]
C
H
.
s
c
[
1
v
2
4
4
3
0
.
4
0
3
2
:
v
i
X
r
a
Generative Agents: Interactive Simulacra of Human Behavior
Joon Sung Park
Stanford University
Stanford, USA
joonspk@stanford.edu
Meredith Ringel Morris
Google Research
Seattle, WA, USA
merrie@google.com
Joseph C. O’Brien
Stanfor... | Generative Agents- Interactive Simulacra of Human Behavior |
38
11 Conclusion
In this survey, we have systematically explored the realm of resource-efficient Large
Language Models (LLMs), offering a comprehensive view of the current state-of-the-art
techniques and methodologies. We started by providing a foundational understanding
of the challenges and necessities in developing ... | Beyond Efficiency |
field regression loss. In contrast to auto-regressive models, Voicebox can consume context not only
in the past but also in the future. Moreover, the number of flow steps can be controlled at inference
time to flexibly trade off quality and runtime efficiency.
Voicebox is trained on 60K hours of English audiobooks and ... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Democratic Creative Destruction?
155 | Social_Media_and_Democracy |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Internet Platforms and Content Moderation
251
West, L. (2017). I’ve left Twitter. It is unusable for anyone but trolls, robots and
dictators. The Guardian, January 3. www.theguardian.com/commentisfree/2017/
jan/03/ive-left-twitter-... | Social_Media_and_Democracy |
Unpublishedworkingdraft.
Notfordistribution.
Median 95% HDI
Table 8. Model outputs for the parameters on the log scale. Medians are provided for each parameter, along
with their 95% HDI and 𝑝𝑏. Parameters distinguishable from zero are marked with *. We ran the model with
two chains and 4000 iterations.
D MODEL PAR... | AI enhance sour performance |
content found within the acquired training datasets, frequently encompassing toxic linguistic
elements, including offensive, hostile, and derogatory language [50].
• LLMs may exhibit social biases and toxicity [35, 50, 144] during the generation process,
resulting in the production of biased outputs.
• LLMs may manifes... | ASurveyonEvaluationofLargeLanguageModels |
(McGinnies and Ward 1980; Guillory and Geraci 2013). People are more
likely to view sources as trustworthy if they share similar traits. As a result,
corrections that are attributed to an in-group member (e.g., a leader of one’s
preferred party) may be more effective than those credited to an out-group
member (e.g., an... | Social_Media_and_Democracy |
Table 2. Team Design Pattern: AI Performer and Human Assistant
The AI performs a certain task (0) and monitors its own state to determine whether
there are aspects that need human involvement (1). Once the AI recognizes that hu-
man intervention is necessary, it proactively requests assistance (2) from the human
by co... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
6
Published as a conference paper at ICLR 2023
Method
AdaRound
AdaQuant
BRECQ
OBQ
GPTQ
RN18 – 69.76 % RN50 – 76.13%
4bit
69.34
68.12
69.37
69.56
69.37
3bit
68.37
59.21
68.47
68.69
67.88
4bit
75.84
74.68
75.88
75.72
75.71
3bit
75.14
64.98
75.32
75.24
74.87
Table 1: Comparison with state-of-the-art
post-training... | GPTQ |
intelligence in education, machine learning, natural language processing, automatic
assessment
Automated assessment (AA), the task of employing machine learning models to automatically
score written/spoken text, is one of the most important educational applications of natural
language processing. Emerged as a mea... | informatics-phd-projects-2022-23 |
Data Undersampling. Beside deduplication, data undersampling, also referred to as instance selection, emerges as another
promising data filtering technique [316]. This approach aims to reduce the volume of training samples by sub-sampling
large datasets, yet crucially retains the distribution characteristics of the ori... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
A Review of Deep Learning Techniques for Speech Processing
105
[529] Chuanxin Tang, Chong Luo, Zhiyuan Zhao, Wenxuan Xie, and Wenjun Zeng. 2021. Joint time-frequency and time
domain learning for speech enhancement. In Proceedings of the Twenty-Ninth International Conference on International
Joint Conferences on Artif... | AReviewofDeepLearningTechniquesforSpeechProcessing |
generates a new program and does not rely on code snippets in the initial prediction. See Figure 11
for an example where the prediction after SELF-DEBUGGING is very different from the initial code. | Teaching Large Language Models to Self-Debug |
Pick up a oak_wood in Forest.
Pick up a oak_slab in Forest.
Pick up a oak_planks in Forest.
Pick up a oak_log in Forest.
Pick up a oak_button in Forest.
Pick up a oak_door in Forest.
Pick up a oak_fence in Forest.
Pick up a oak_fence_gate in Forest.
Pick up a oak_trapdoor in Forest.
Pick up a oak_boat in Forest.
Pick u... | JARVIS-1 |
Routing Strategy Routing strategy is an essential component of Mixture-of-Experts (MoE) models,
playing a pivotal role in determining the effectiveness and efficiency of these models. The primary
function of the routing strategy is to intelligently distribute input data among multiple specialized
experts, each optimize... | Mixture-of-Experts |
For summarization in the low data settings, we
use a learning rate of 5e-5 and a warmup step of
100. We use a batch size of 5 for prefix-tuning
and 6 for fine-tuning. We apply the initialization
trick and use the word “summarize” to initialize
4594 A.4 Additional Results for Low-data Settings
Figure 7 supplements the lo... | Prefix-Tuning |
n
d
a
t
a
:
t
h
e
h
u
m
a
n
e
x
p
l
a
n
a
t
i
o
n
b
a
s
e
l
i
n
e
,
a
n
d
h
u
m
a
n
s
c
o
r
i
n
g
e
x
p
e
r
i
m
e
n
t
s
.
N
i
c
k
a
n
d
W
i
l
l
i
a
m
i
m
p
l
e
m
e
n
t
e
d
e
a
r
l
y
r
e
s
e
a
r
c
h
e
r
e
x
p
l
a
n
a
t
i
o
n
b
a
s
e
l
i
n
e
s
.
A
l
t
e
r
n
a
t
i
v
e
t
o
k
e
n
a
n
d
... | Language models can explain neurons in language models |
access to social media data to third parties for research purposes; but
3. There are real differences between private actors who analyze these data in
order to support for-profit businesses with no obligation to release find-
ings to the public (and indeed may even have obligations to shareholders
not to do so) and othe... | Social_Media_and_Democracy |
Captions [5, 27], and SBU [20] to align visual features with the Vicuna language model. However,
simply aligning the visual features with the LLM is insufficient to train high-performing model
with visual conversation abilities like a chatbot, and the noises underlying the raw image-text pairs
may result in incoherent l... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
responsive to this fact.
Of course, this is a very simple, simulated environment, and the level of agentic planning it makes
sense to ascribe to these AIs isn’t clear.74 But the basic dynamic that gives rise to this type of behavior
seems likely to apply in much more complex, real-world contexts, and to more sophistica... | Is Power-Seeking AI an Existential Risk? |
Table 7 compares real and synthetic data from Voicebox and three baseline models. Each TTS model
generates one sample per text from the Librispeech training set, resulting in 281K utterances per
system. For real data, we consider train-960 and train-clean-100. Details about the ASR model and
training configurations are... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
8https://www.ncbi.nlm.nih.gov/research/bionlp/Data/
27
MeQSum, iCliniq, HealthCareMagic (Abacha & Demner-Fushman, 2019; Zeng et al., 2020) are abstrac-
tive summarization datasets and we preprocess them following BioBART (Yuan et al., 2022). Specifically,
MeQSum contains 1000 refined patients’ health questions selec... | BiomedGPT |
F (C(p)) : R3 (cid:55)→ [0, 1]
C(p) = (S (FI , π(p)) , Z(p))T
where FI = EI (I) represents the image feature map from
the deep image encoder EI (·), π(p) the 2D projection of
p on the feature map FI, S (·,·) is the sampling function
used to sample the value of FI at pixel π(p) using bilinear
interpolation, and Z(p) i... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
Our model learns representations for a pre-fixed
vocabulary of entities, and cannot handle unseen
entities. Future work can explore representations
for rare or unseen entities, as well as developing
less memory-intensive ways to learn and integrate
entity representations. Furthermore, integrating
information from knowle... | Entities as Experts- Sparse Memory Access with Entity Supervision |
Training a Helpful and Harmless Assistant with
Reinforcement Learning from Human Feedback
Yuntao Bai∗, Andy Jones, Kamal Ndousse,
Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort,
Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion,
Tom Conerly, Sheer El-Showk, Nelson El... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Chan, L., Garriga-Alonso, A., Goldowsky-Dill, N., Green-
blatt, R., Nitishinskaya, J., Radhakrishnan, A., Shlegeris,
B., and Thomas, N. Causal scrubbing: a method for
rigorously testing interpretability hypotheses. Alignment
Forum, 2022. URL https://www.alignmentfor
um.org/posts/JvZhhzycHu2Yd57RN/causa
l-scrubbing-a-me... | Eight Things to Know about Large Language Models |
to bed.
to bed.
straight to
bed.
straight to
bed.
straight to
bed.
outside
a rainbow!
many
so
colors.
many
colors!
many
colors.
appeared!
appeared.
It’s
pretty.
so
red,
ange,
yellow,
green,
blue,
purple!
or-
and
different.
the
small.
like
moon.
small
round.
and
small
thin.
and
small.
small
thi... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
b
a
s
i
s
a
t
a
l
l
.
[
↩
]
| Language models can explain neurons in language models |
1) PLMs and Datasets: We use the encoder-only models
RoBERTa-base (125M) and RoBERTa-large (355M) [2] to
evaluate on the GLUE benchmark [100], encoder-decoder
models T5-base (220M) and T5-large (770M) [4] to evaluate
on the WMT16 En-Ro dataset6, and decoder-only models
LLaMA-7B and LLaMA-13B [7] fine-tuned with the Alp... | Parameter-EfficientFine-TuningMethods |
Text tokenizer and embedding as input We find that a
strong text encoding is important for accurate and high qual-
ity text-to-video generation. Pretrained text representations
in general outperformed training our model with text tokens
from scratch. Due to computational constraints, we found it
more efficient to leverag... | VideoPoet |
tree/main/projects/OPT/chronicles
GPU Type
GPU Power
consumption
GPU-hours
OPT-175B
A100-80GB
BLOOM-175B A100-80GB
A100-80GB
LLaMA-7B
A100-80GB
LLaMA-13B
LLaMA-33B
A100-80GB
A100-80GB
LLaMA-65B
400W
400W
400W
400W
400W
400W
809,472
1,082,880
82,432
135,168
530,432
1,022,362
(tCO2eq)
Total power Carbon emitted... | LLaMA- Open and Efficient Foundation Language Models |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.