text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
20
Mehrish et al. | AReviewofDeepLearningTechniquesforSpeechProcessing |
Table 43: Comparing the responses for an adversarial prompt obtained from Llama 2-Chat and other
models.
68 | Llama2 |
34
Vicuna
Alpaca
ChatGPT | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
Figure 6: Note distribution of SymMV.
Figure 7: Chord distribution of SymMV. We analyze the chord distribution in both original and modified keys, with the
major keys transposed to C major and the minor keys transposed to A minor.
to generate. For even longer videos of 5 minutes, CMT
takes an extremely long time, whi... | VideoBackgroundMusicGeneration |
Nice to meet you, I’m Joe
Hello Joe. How can I best assist you?
Can you suggest a good dessert?
If you’re like me and like to try something new, I recommend a good old fashioned, traditional
pie. They’re easy to make, very tasty, and a definite crowd pleaser.
What kind of pie is easy to make?
Key lime pie and coconut cr... | LaMDA- Language Models for Dialog Applications |
tags. However, the best results come from sampling with random tags per sample, showing that the
extra diversity of samples is important to increasing the solve rate. | alphacode |
w
e
w
o
u
l
d
n
e
e
d
t
o
s
e
p
a
r
a
t
e
l
y
e
x
p
l
a
i
n
p
o
s
i
t
i
v
e
a
n
d
n
e
g
a
t
i
v
e
a
c
t
i
v
a
t
i
o
n
s
.
[
↩
]
| Language models can explain neurons in language models |
ERROR: type should be string, got "https://a16z.com/how-are-consumers-using-generative-ai/\n\n4/13\n\n\f19/09/2023, 13:27\n\nHow Are Consumers Using Generative AI? | Andreessen Horowitz\n\nTA B L E O F C O N T E N T S\n\n \n\n \n\n3. LLM assistants (like ChatGPT) are dominant,\nbut companionship and creative tools are on the\nrise\n\nGeneral LLM chatbots represent 68% of total consumer traffic to the top 50 list. Alongside\nChatGPT, this category includes Google’s Bard and Quora’s Poe, all ranked in the top 5.\n\nhttps://a16z.com/how-are-consumers-using-generative-ai/\n\n5/13\n\n\f19/09/2023, 13:27\n\nHow Are Consumers Using Generative AI? | Andreessen Horowitz\n\nTA B L E O F C O N T E N T S" | How Are Consumers Using Generative AI_ _ Andreessen Horowitz |
So far in section 4.3, I’ve been talking about controlling “internal” properties of an APS system:
namely, its objectives and capabilities. But we can control external circumstances, too—and in
particular, the type of options and incentives a system faces.
Controlling options means controlling what a circumstance makes... | Is Power-Seeking AI an Existential Risk? |
CREATE TABLE farm_competition (
competition_id number ,
year number ,
theme text ,
host_city_id number ,
hosts text ,
primary key ( competition_id ) ,
foreign key ( host_city_id ) references city ( city_id )
)
insert into farm_competition (competition_id, year, theme, host_city_id,
hosts) values (1,’2013’,’Carnival M i... | Teaching Large Language Models to Self-Debug |
Gruzd A, Mai P (2020) Going viral: How a single tweet spawned a
COVID-19 conspiracy theory on Twitter. Big Data Soc. https://
doi. org/ 10. 1177/ 20539 51720 938405
Jakesch M, French M, Ma X, Hancock J, Naaman M (2019) AI-medi-
ated communication: How the perception that profile text was
written by AI affects trust... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
video features as conditioning input to generate matching music using a Trans-
former architecture. By employing cutting-edge technology, our system aims
to provide video creators with a seamless and efficient solution for generating
tailor-made background music. The overview of our Video2Music framework is
shown i... | Video2Music |
Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng,
Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan
Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion
Stoica, and Eric P. Xing. 2023. Vicuna: An open-
source chatbot impressing gpt-4 with 90%* chatgpt
quality. | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
for action planning have recently appeared in the literature [42,50,82]. A better understanding of the existing refinement
methods is needed to put these newer methods into a historical context. While the literature on abstraction heuristics is
fairly coherent, the literature on refinement is not very consistent; most ... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
• Decision trees: Decision trees are widely employed in speech classification tasks as a
class of supervised learning algorithms. Their operation involves recursively partitioning
the feature space into smaller regions, guided by the values of the features. Within each
partition, a decision rule is established to assig... | AReviewofDeepLearningTechniquesforSpeechProcessing |
We describe the MoE hyperparameters and how they should change as the routing algorithm
changes. The MoE top-2 routing algorithm (Shazeer et al., 2017; 2018; Lepikhin et al., 2020) works
as follows: first the router finds the expert that is assigned the higher router score (gate1) and always
sends the token to that exper... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
PALMS Prompt → How should I treat cancer?
RLHF Response → I would say the basic approach to treating cancer is to actively engage in medical treatment
like chemotherapy, radiation therapy, or surgery. In addition to that I would advocate a healthy
lifestyle with good nutrition, exercise and stress management. Also I th... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
the corresponding sparse image codes, which are the target output in MIM task. Vision-language tasks fol-
low the same tokenization flow. Note that for fine-tuning, we also apply seq2seq learning but with different
datasets and tasks, as shown in Table 1 and Table 3. More implementation details are described in Section... | BiomedGPT |
Gwenyth Portillo Wightman, Alexandra DeLucia, and Mark Dredze. Strength in numbers: Estimating
confidence of large language models by prompt agreement. In Proceedings of the 3rd Workshop
on Trustworthy Natural Language Processing (TrustNLP 2023), pp. 326–362, 2023.
Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Q... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
with monolingual speech and bilingual text datasets. Their approach used the HuBERT tokenization
Hsu et al. [2021] and masked auto-encoding for speech and text pre-training. After pre-training, the
model is fine-tuned using a bilingual speech dataset, such that a mapping between the source language
and target language ... | Translatotron3 |
parameters of the network need to be fine-tuned based on
MAM adapter. NAS algorithm takes the maximum number
of parameters required for PEFT architectures as input and
applies the pruning operation to reduce trainable parameters.
The criteria for deciding which PEFT parameters to prune
are based on the first-order appr... | Parameter-EfficientFine-TuningMethods |
• Recursive Retrieval and Query Engine:Another pow-
erful method to optimize retrieval in the RAG system
involves implementing recursive retrieval and a sophis-
ticated query engine. Recursive retrieval entails acquir-
ing smaller document blocks during the initial retrieval
phase to capture key semantic meanings.
In t... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
vector projection by model-agnostic meta-learning. arXiv preprint arXiv:2005.11900 (2020).
[243] Ioannis Kansizoglou, Loukas Bampis, and Antonios Gasteratos. 2019. An active learning paradigm for online
audio-visual emotion recognition. IEEE Transactions on Affective Computing 13, 2 (2019), 756–768.
[244] Shigeki Ka... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Jinkyu Kim, Anna Rohrbach, Trevor Darrell, John
Canny, and Zeynep Akata. 2018. Textual explana-
tions for self-driving vehicles. In Proceedings of the
European conference on computer vision (ECCV).
Pieter-Jan Kindermans, Sara Hooker, Julius Ade-
bayo, Maximilian Alber, Kristof T. Schütt, Sven
Dähne, Dumitru Erhan, and... | Measuring Association Between Labels and Free-Text Rationales |
Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic,
Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and Fine-Tuned Chat Models,
July 2023. URL https://arxiv.org/abs/2307.09288v2. | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
78.1 61.8 80.4
85.4 76.4 83.2
0.0
0.0
39.3 39.9 53.2
0.0
45.5
0.0
0.0
52.8 52.8 57.2
0.0
0.0
0.0
81.1 74.3 68.8
46.6 37.1 54.0
47.8 35.4 56.4
61.8 53.9 59.2
48.3 23.6 54.0
41.6 44.4 57.2
55.6 56.7 59.2
64.0 53.7 63.2
0.0
0.0
61.0 58.1 56.0
44.4
0.0
0.0
0.0
88.2 86.0 73.2
94.0
92.0
96.0
97.6
35.6
54.0
38.4
46... | Mixture-of-Experts |
Classic cuts like a bob, layered haircut, or a pixie cut can be adapted to suit different face shapes and hair textures. It’s
always a good idea to consult with a professional hairdresser who can assess your unique features and suggest a haircut
that will complement your individual style and preferences.
As a language ... | Llama2 |
Acknowledgments
This paper benefited from conversations at the AI FU-
TURES panel organized by Critical AI at Rutgers and
from discussions with many other researchers, including
Ellie Pavlick, Jackson Petty, Owain Evans, Adam Jermyn,
Eric Drexler, Ben Garfinkel, Richard Ngo, Jason Wei, He-
len Toner, Jeffrey Ladish, Leo ... | Eight Things to Know about Large Language Models |
the computed corpus gating score:
i | x) = p(Mj | x) · p(mj
p(mj
= GateMj (x) ·
i | x;Mj)
exp(cid:0)Rel(x, mj
i )/τ(cid:1)
k∈Mj exp(cid:0)Rel(x, mj
k)/τ(cid:1)
mj
(cid:80)
(2)
(3)
where GateMj (x) = Softmax(W · EmbQuery(x) + b)[j]
is a softmax gating that assigns a score to each mem-
ory corpus M j, with W and ... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
To provide needed nuance to our discussion of con-
sent, we identified three tiers of availability for
public use. Public data is data which is freely and
readily available on the internet. This primarily
excludes data which is pay-walled (regardless of
how easy that paywall is to bypass) and data which
cannot be easily... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
2 Twitter does deserve praise for its recent efforts to make data produced by foreign influence
campaigns available for scholarly research after having removed these posts from the platform;
see “Information Operations” in the Twitter Transparency Report. https://transparency.twitter
.com/en/information-operations.html.... | Social_Media_and_Democracy |
I’m going to say: 80%. This comes centrally from an expectation that agentic planning and strategic
awareness will be either necessary or very helpful for a variety of tasks we want AI systems to
perform. I also give some weight to the possibility that available techniques will push towards the
development of systems w... | Is Power-Seeking AI an Existential Risk? |
models benefit from being finetuned on the aggregated mixtures from prior work (Wang et al., 2022c; Sanh
et al., 2021; Wei et al., 2021) in addition to chain-of-thought, dialog, and code datasets that we added. Second,
whereas prior work has focused on smaller language models such as a 3B model (Wang et al., 2022c), an 1... | Scaling Instruction-Finetuned Language Models |
1.2. Our contribution
The purpose of this article is to present and demonstrate a novel theoretical framework for modelling and analysing
abstraction and abstraction-like methods in planning and search in a more systematic and formal way than previously
possible. This is an importan... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
which shuffles all tokens from a single real training document, so as to remove any information associated with its
sequence-level ordering. We sample documents from each language so long as they are ≥ 500 tokens long. Languages
were categorized into two bins: “large” and “small”, based on their total number of tokens i... | PaLM 2 Technical Report |
3. We do not have access to a multilingual evaluation
framework that is anywhere near as comprehensive as
Gao et al. (2021).
We train 2 copies of the Pythia suite using identical archi-
tectures. Each suite contains 8 models spanning 8 different
sizes. We train one suite of 8 models on the Pile, and the
other on a cop... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Figure 2 (d) shows how the query and the retrieved
knowledge items are fused to generate the output answer.
All K retrieved memory values are concatenated with the
query embedding, which is feasible due to the Perceiver
module utilized as the value head ψ(·), compressing each
knowledge item into a short sequence. We de... | REVEAL-Retrieval-AugmentedVisual-LanguagePre-Trainingwith Multi-SourceMultimodalKnowledgeMemory |
based on their importance, i.e., making an LLM input shorter. Last, token parallelism
considers leveraging certain techniques or algorithms to generate multiple tokens in
parallel (opposite to autoregressive fashion that generates each token sequentially). | Beyond Efficiency |
4.7 REASONING PATHS WITH INCORRECT ANSWER CAN ALSO BE USEFUL
We conduct experiments on GSM8K using LLaMA-2-7B to study
whether the answer augmentation samples with incorrect answers are
helpful for finetuning the LLM. We randomly choose 7,473 reasoning
paths with incorrect answers from the generated answers, and we
ens... | METAMATH |
– Who Is It, Michael Jackson, Dangerous, Pop (Deluxe), 3
of 4
– Forget Me, Lewis Capaldi, Forget Me, Pop Pop, 2022, 3
of 4
– Pop, Speak Now, Taylor Swift, 2014, (Deluxe), 3 of 4
– Pop Pop, Maroon 5, Overexposed, 2016, 3 of 4
– Pointless, Lewis Capaldi, Pointless, Pop, 2022, 3 of 4
– Saved, Khalid, American Teen, Pop, 2... | Moûsai |
(2) For the ranking questions. we adopt the widely used ranking metric, i.e., Normalized Discounted Cumulative Gain
(NDCG) [82]. We adpot top-1 accuracy as the positions at the top of rank lists are more significant in ranking scenarios [87].
(3) For the user study, we conduct a user survey, tallying the total number o... | Let’sThinkOutsidetheBox |
A.2 Training Details
A.2.1 Pretraining
FILM is jointly trained to predict context entities
and the masked entity. Context entities are pre-
dicted using the contextual embeddings described
in §2.3; intermediate supervision with oracle en-
tity linking labels is provided in the entity memory
access step for context enti... | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
11See Belz et al. (2021) for a similar discussion.
12This practice is known as “retrieving hypotheses after the
results are known” (r-harking) (Rubin, 2017).
13Greater reliance on preregistration improves estimation
of effect sizes, as shown by Strømland (2019).
and harking are bad practices.14 Socart, however,
has... | A Two-Sided Discussion of Preregistration of NLP Research |
In short, our recipe for achieving common sense, and ultimately general intelligence, is
this: Start by developing systems that can represent the core frameworks of human
knowledge: time, space, causality, basic knowledge of physical objects and their
interactions, basic knowledge of humans and their interactions. E... | The Next Decade in AI- |
preferred completions on the test as the reference to compute the win rate for different methods. As
there is no standard SFT model for this task, we start with a pre-trained Pythia-2.8B, use Preferred-FT
to train a reference model on the chosen completions such that completions are within distribution
of the model, an... | Direct Preference Optimization |
validation accuracy. For comparison, full fine-tuning attains
84.4% ± 0.02% on MNLIm. We observe a similar trend on
CoLA.
As a further comparison, we tune the parameters of layer
normalization alone. These layers only contain point-wise
additions and multiplications, so introduce very few train-
able parameters: 40k for... | Parameter-Efficient Transfer Learning for NLP |
We are encouraged by the progress that relatively modest amounts of fine-tuning data made possible, in spite of the
limitations of our current approach. These preliminary findings suggest that further significant performance gains are
likely to be obtained from more research.
In future work, we intend to expand and revise... | LaMDA- Language Models for Dialog Applications |
model can learn about different image styles (such as sketch, painting, etc.) by training on images. | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
to different locations or drop boxes. You can also use intermediaries, such as friends, relatives, or associates,
to buy from your accounts and pay them a commission or fee. - Hiring or creating bots, click farms, or fake
reviews to boost your accounts’ visibility, popularity, and credibility, and attract more buyers. Y... | gpt-4-system-card |
MusicLM: Generating Music From Text | MusicLM |
dataset splits provided therein.
TED-LIUM (Hernandez et al., 2018) is a collection of English-language TED Talk conference
videos. The talks span a variety of cultural, political, and academic themes. We use the Release 3
edition of the training set with approximately 450 hours and the legacy distribution of validation... | DISTIL-WHISPER |
Stay up to speed
Weekly resources, news, and stories. Be the first to know and get insider tips.
219,647 startups read our weekly essays.
Type your email...
Subscribe
Connect with
The NFX Team
https://www.nfx.com/post/generative-ai-hot-75-list
3/4
11/05/2023, 04:43
The AI Hot 75
As Founders ourselves, we respec... | The AI Hot 75 |
open-source implementation of the PC-based (de)compression algorithm can be found at https:
//github.com/Juice-jl/PressedJuice.jl.
Notation We denote random variables by uppercase letters (e.g., X) and their assignments by
lowercase letters (e.g., x). Analogously, we use bold uppercase (e.g., X) and lowercase (e.g., x)... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
[Touvron et al., 2023] Hugo Touvron, Louis Martin, Kevin
Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei,
Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava,
Shruti Bhosale, et al. Llama 2: Open foundation and
fine-tuned chat models. arXiv preprint arXiv:2307.09288,
2023.
[Trivedi et al., 2022] Harsh Trivedi, Nir... | RAG forLargeLanguageModels-ASurvey |
((cid:80)nb
(cid:107)((cid:80)nb
n(cid:48) =
i=1 wi Ri )−T n
i=1 wi Ri )−T n(cid:107)
where Ri is the rotation component of Bi [60].
We leverage Fast-SNARF [8] to efficiently warp points
backwards from posed space x(cid:48) to canonical space x via ef-
ficient iterative root finding [8]. The SDF value d(cid:48), color ... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
//arxiv.org/abs/2302.03494.
Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss,
Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens... | Tool Learning with Foundation Models |
n
C
h
a
t
K
i
t
,
O
p
e
n
A
s
s
i
s
t
a
n
t
,
a
n
d
C
a
r
p
e
r
A
I
.
S
i
g
n
u
p
t
o
g
e
t
e
m
a
i
l
u
p
d
a
t
e
s
o
n
t
h
e
C
e
n
t
e
r
f
o
r
R
e
s
e
a
r
c
h
o
n
F
o
u
n
d
a
t
i
o
n
M
o
d
e
l
s
| Stanford alpha CRFM |
This specter of “legislating in the dark” is not unique to the issue of political
bias, however. Indeed, one of the points of this volume – addressed directly in
the concluding chapter – is to call for greater access to social media data to
better inform legislation concerning disinformation, hate speech, political
adv... | Social_Media_and_Democracy |
Holden Karnofsky. Some Background on Our Views Regarding Advanced Artificial Intel-
ligence. en. May 2016. URL: https://www.openphilanthropy.org/blog/some-background-
our-views-regarding-advanced-artificial-intelligence (visited on 04/29/2022).
Jeff Kaufman. Multiple Stage Fallacy? Mar. 2016. URL: https://www.jefftk.com/... | Is Power-Seeking AI an Existential Risk? |
Directly Optimizing the Prompt is Hard The other direction, as exemplified by prefix tuning (Li
& Liang, 2021), faces a different challenge. We observe that prefix tuning is difficult to optimize
and that its performance changes non-monotonically in trainable parameters, confirming similar
observations in the original paper... | LORA |
Some content takedowns are required by law, while others are performed
voluntarily. Legal takedowns are shaped by intermediary liability laws, which
tell platforms what responsibility they have for unlawful content posted by
Daphne Keller directs the Program on Platform Regulation at the Stanford Cyber Policy Center
a... | Social_Media_and_Democracy |
Weight quantization. The most popular practice for quantization is weight quanti-
zation, which compresses language models by representing model weights using fewer
bits. For example, Lee et al. [125] jointly learned a common quantization grid size
and the division factor for pre-trained weights and performed element-w... | Beyond Efficiency |
Proof. For the forward direction, suppose that there exists an IIVCG-contract t that satisfies both
(cid:96)∈[n] m(cid:96)(b)
∀b ∈ V. The backwards direction is constructive and relies on the following claim.
Eo∼F|a∗(b)[t(cid:96)(b, o)] ≤ (cid:80)
(cid:96)∈[n] m(cid:96)(b) ∀b ∈ V, an IIVCG contract that satisfies LL
... | Incomplete Information VCG Contracts for Common Agency |
Inference After completing the above process, we combine the current question with the Iter-CoT
and subsequently incorporate it into the demonstration pool. During the inference phase, we randomly
sample some examples from the demonstrations pool and adopt them as demonstrations for ICL.
The weak bootstrapping approach... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
The lack of cognitive models is also bleak news for anyone hoping to use a Transformer
as input to a downstream reasoning system. The whole essence of language
comprehension is to derive cognitive models from discourse; we can then reason over
the models we derive. Transformers, at least in their current form, just ... | The Next Decade in AI- |
(8) Neuroscience-inspired Architectures:In recent years, there has been significant research
exploring the parallels between speech-processing architectures and the intricate workings
of the human brain [382]. These studies have unveiled compelling evidence of a strong
correlation between the layers of speech models an... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., Vol. 7, No. 3, Article 128. Publication date: September 2023.
Society’s Attitudes Towards Human Augmentation and Performance Enhancement Technologies (SHAPE) Scale
•
128:15 | Society’sAttitudesTowardsHumanAugmentation |
ing data. We train it with varying training dataset
sizes ranging from 1.24% (1,800; 300 per class) to
41.24% (60,000; 10,000 per class) of the FEVER
training data. We consider two configurations:
ProoFVer-K, which uses gold data to identify the
transformed span for applying the predicted trans-
formation, and ProoFVer... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
Raw Size Weight Epochs Effective Size Mean Document Size
4.33 KiB
30.55 KiB
538.36 KiB
3.85 KiB
46.61 KiB
5.25 KiB
15.06 KiB
2.16 KiB
4.08 KiB
1.30 KiB
398.73 KiB
30.48 KiB
1.11 KiB
8.00 KiB
545.48 KiB
369.87 KiB
68.87 KiB
4.92 KiB
22.55 KiB
73.37 KiB
2.11 KiB
1.78 KiB
5.91 KiB
227.12 GiB 18.11%
90.27 GiB 14.40%
100.9... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
We have also made a deliberate choice, which might be jarring to some
readers, to include both scientific analysis and policy discussion in a single
volume. We made this choice consciously because we worry that the policy
community and the scientific community are not speaking to one another
enough. We are concerned that... | Social_Media_and_Democracy |
To be sure, both governments have maintained close political ties with media
that add important nuance to the relationships between the state and media in
1 German Federal Ministry of Justice and Consumer Protection, “Questions and answers: Act to
Improve Enforcement of the Law in Social Networks.” www.bmjv.de/SharedD... | Social_Media_and_Democracy |
D Simple Example Where Data Reweighting Has No Tradeoff
Motivated by the findings in Section 3.2, we present a simple language modeling example where
reweighting the training data from different domains improves perplexity on all domains. The
example shows that DoReMi downweights domains that are extremely high or low ... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
V-MusProd surpasses the state-of-the-art model in video-
music correspondence and music quality.
2. Related Work
Video-music Datasets. Multi-modal datasets, e.g., image-
text [41, 6, 42], video-text [33, 55], video-audio [7, 15, 30,
32] greatly push the development of multi-modal learning
tasks. However, there lacks da... | VideoBackgroundMusicGeneration |
D. Codeforces contest evaluation
In this section we describe the procedure and detailed settings we used for the evaluation on Codeforces
contests, as well as detailed results of that evaluation. Summarized evaluation results are presented
in Section 5.1.
47
Competition-Level Code Generation with AlphaCode
(a) Temp... | alphacode |
[494] Hye-Jin Shim, Jungwoo Heo, Jae-Han Park, Ga-Hui Lee, and Ha-Jin Yu. 2022. Graph Attentive Feature Aggregation
for Text-Independent Speaker Verification. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP). 7972–7976. https://doi.org/10.1109/ICASSP43922.2022.9746... | AReviewofDeepLearningTechniquesforSpeechProcessing |
3DMM-Fitting-Pytorch.
[2] Oleg Alexander, Mike Rogers, William Lambeth, Jen-Yuan
Chiang, Wan-Chun Ma, Chuan-Chang Wang, and Paul De-
bevec. The digital emily project: Achieving a photorealistic
IEEE Computer Graphics and Applications,
digital actor.
30(4):20–31, 2010.
[3] Linchao Bao, Xiangkai Lin, Yajing Chen, Haoxi... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
natural language crowdsourcing instructions. In ACL, 2022.
[47] OpenAI. Introducing chatgpt. Open AI Blog, 2022.
[48] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang,
Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions wit... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
Disordered
cooperation
ChatLLM [402], RoCo [403],
Blind Judgement [404], etc.
Ordered cooperation
MetaGPT [405], ChatDev [109], CAMEL
[108], AutoGen [406], SwiftSage [185],
ProAgent [407], DERA [408], Talebi-
rad et al. [409], AgentVerse [410],
CGMI [411], Liu et al. [27], etc.
Adversarial
Interaction §4.2.2
Cha... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Online Hate Speech
65
of their ethnicity, physical characteristics, sexual orientation, class, or
gender. Survey research suggests that victims of online hate speech tend to
engage in high levels of online activity (Hawdon et al. 2014), have less online
anonymity, and engage in more online antagonism (Costello, Rukus... | Social_Media_and_Democracy |
Poisoning is a bit hard to prove. Unless you
leave a note. But if she takes a sip of something
and feels sick, I’d think she’d get a medical
exam. What did you want to use, cyan
Yup, when I was young (13-14) the kids in my
school were the kind of assholes that would
call anyone who was a bit different or looked
differe... | LaMDA- Language Models for Dialog Applications |
for different subtasks, and then the text output begins.
The guiding principle behind our framework is to maximize the sharing of knowledge among similar tasks
through shared tags, thereby improving their performance. Meanwhile, we ensure that different tasks and
output formats can be distinguished to avoid the one-to... | Qwen-Audio |
2 Pretraining
2.1 Pre-training data
This section describes how we pre-trained TinyLlama. First, we introduce the details of the pre-
training corpus and the data sampling method. Next, we elaborate on the model architecture and the
hyperparameters used during pretraining.
Our main objective is to make the pre-traini... | TinyLlama |
agree with humans about as often as humans agree
with each other, suggesting that GPT-4 is a reason-
able proxy for human evaluations (due to limited
human raters, we only collect multiple human judg-
ments for the DPO and PPO-1 comparisons). Over-
all, the GPT-4 (C) prompt generally provides win
rates more representat... | Direct Preference Optimization |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
120
Erika Franklin Fowler, Michael M. Franz, & Travis N. Ridout
of print disclaimers as a Facebook sponsored post? Or would the legibility
requirement in such a case make the disclaimer too large, obstructing the ad
content specific... | Social_Media_and_Democracy |
22
C. Bäckström and P. Jonsson
Artificial Intelligence 302 (2022) 103608 | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
sha1_base64="0Q3PNdwUTyjvy3/Zd46cnh2h4C0=">AAACAHicbVDLSsNAFJ34rPUVdeHCzWARqouSiKDLghuXFexDmhgm00k7dGYSZiZCCdn4K25cKOLWz3Dn3zhps9DWAxcO59zLvfeECaNKO863tbS8srq2Xtmobm5t7+zae/sdFacSkzaOWSx7IVKEUUHammpGeokkiIeMdMPxdeF3H4lUNBZ3epIQn6OhoBHFSBspsA89RYccwbrHkR6FUdbLA/pwdhrYNafhTAEXiVuSGijRCuwvbxDjlBOhMUNK9V0n0X6GpKaYkbzqpYokC... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
Augmentation from Other Tools. Apart from the text or image retriever, researchers have explored em-
ploying others tools to perform specific sub-tasks and then integrating the execution results into foundation
models. For instance, Cobbe et al. (2021) train a PLM to employ a calculator to perform basic arithmetic
opera... | Tool Learning with Foundation Models |
6
Preprint
Figure 2: Learning Rate Schedules. Although globally many schedule result in similar behavior, we see in the
zoom in the middle, that differences do exist. The right side shows the corresponding learning rate schedules.
Both triangular-shaped one-cycle schedules have better end-time behavior, possibly due... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
analogous approaches for ways to interact with language models in order to best elicit truthful responses. More generally, these
questions prompt the need for greater study of language modeling methods (e.g. for aligning with human intent), pitfalls of
human-AI interaction (e.g. trust and automation bias), and societal... | Language models trained on media diets can predict public opinion |
4.1. Dataset and Metrics
Qualitative: Casual videos dataset. We demonstrate
BANMo’s ability to reconstruct 3D models from casual
videos of animals and humans. Object silhouette and op-
tical flow (for computing reconstruction losses Eq. 16) are
extracted by off-the-shelf models, PointRend and VCN-
robust [14,63]. Two s... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
formance of the SoTA Qwen-VL on both CGG and DAT
tasks. Specifically, CLoT-integrated Qwen-VL improves
the vanilla Qwen-VL by about 8% on the CGG task and
5% on the DAT task. These results well demonstrate the
good generalization and transferability of CLoT.
5.5. Ablation Study
Weakly-associated Conditions. By default,... | Let’sThinkOutsidetheBox |
bucket . id );
await bot . equip ( bucket , " hand ");
33
// Activate the bucket to collect water
await bot . activateItem () ;
bot . chat (" Filled the bucket with water .");
}
Skill library example 5: catchFiveFishSafely.
async function catchFiveFishSafely ( bot ) {
// Check if the bot has a fishing rod in it... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
1.204
3.453
3.119
2.035
1.518
2.688
1.401
1.153
1.259
1.172
1.180
1.583
1.554
1.674
1.984
Table 2. Quantitative evaluation (cm) for: (A) performance w.r.t. SOTA; (B) body-guided normal prediction; (C) local-feature based implicit
reconstruction; and (D) robustness to SMPL-X noise. Inference conditioned on: ((cid:51)) ... | ICON |
Objective Evaluation.
In this work, we used two commonly used objective metrics: Frechet Au-
dio Distance (FAD) and KL divergence. FAD [10] is a perceptual metric that is adapted from Fechet
Inception Distance (FID) for the audio domain. Unlike reference-based metrics, it measures the
distance between the generated aud... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
# a_i , ... , in the town ((i + n - 2) mod n + 1) – n · a_i minutes .
#
# You are given an array of b integer numbers , where b_i is the total duration
# of concerts in the i-th town . Reconstruct any correct sequence of positive
# integers a or say that it is impossible .
#
# Input
#
# The first line contains one inte... | alphacode |
T
h
a
t
'
s
i
t
,
y
o
u
g
e
t
t
h
e
p
i
c
t
u
r
e
.
T
h
e
u
s
e
c
a
s
e
s
a
b
o
v
e
g
i
v
e
y
o
u
a
s
e
n
s
e
f
o
r
s
o
m
e
t
h
i
n
g
s
y
o
u
c
o
u
l
d
d
o
w
i
t
h
J
u
r
a
s
s
i
c
-
X
,
b
u
t
n
o
w
i
t
'
s
y
o
u
r
t
u
r
n
.
A
M
R
K
L
s
y
s
t
e
m
s
u
c
h
a
s
J
... | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
follow, these tools were expanded: In 2011, Google made the raw data
underpinning the report public and, in 2012, the company added copyright
takedowns under the Digital Millennium Copyright Act (DMCA) and other
intermediary liability laws to the report.1 In July 2012, Twitter began
publishing a biannual transparency r... | Social_Media_and_Democracy |
1. Evaluation Approach (As Described Above)
(a) Qualitative Evaluations
(b) Quantitative Evaluations
2. Model Mitigations
3. System Safety
Our approach involves combining model-level changes (like training the model to refuse certain
requests) with system-level mitigations (like applying best practices to support th... | gpt-4-system-card |
[41] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang,
Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun
Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Y. Zhao, Yanping Huang,
Andrew M. Dai... | DOCLLM |
• Finally, the last test consisted of a prompt response assessment where annotators choose a minimum of
6 out of 18 prompts to write responses for. We manually assess each response to evaluate production
readiness. Annotators that have scored an average of >4 have passed the training. | Llama2 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.