text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
The latest technological innovations are constantly being harnessed for
possible transparency initiatives – the recent spread of projects using
distributed ledger systems (e.g., blockchain) to create open registries and
databases provides perhaps the best example (Underwood 2016). Bentham
would likely have approved of ... | Social_Media_and_Democracy |
Scao, T. L., Fan, A., Akiki, C., Pavlick, E., Ili´c, S., Hesslow,
D., Castagn´e, R., Luccioni, A. S., Yvon, F., Gall´e, M.,
Tow, J., Rush, A. M., Biderman, S., Webson, A., Am-
manamanchi, P. S., Wang, T., Sagot, B., Muennighoff, N.,
del Moral, A. V., Ruwase, O., Bawden, R., Bekman, S.,
McMillan-Major, A., Beltagy, I., ... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
PROMPT FOR MATH WORD PROBLEMS
Q: There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done, there
will be 21 trees. How many trees did the grove workers plant today?
A: There are 21 trees now and there are 15 trees in the beginning, so the workers plant 21 - 15 = 6 trees. T... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Of course, we are not the first to notice these compelling advantages. Some of the leading approaches
for leveraging frozen models include prompt tuning (Lester et al., 2021), prefix tuning (Li & Liang,
2021), adapter tuning (Rebuffi et al., 2017; Houlsby et al., 2019), and low rank adaptation (Hu et al.,
2021). All of th... | STANDING ON THE SHOULDERS OF GIANT FROZEN LANGUAGE MODELS |
7
0000Predictor’s Output Down Proj
RowsUp Proj
ColumnsFlash Memoryload
from
flash from Flash Memory. When data is read from flash
memory, the operating system typically caches
these pages, anticipating future reuse. However,
this caching mechanism consumes additional mem-
ory in DRAM beyond what is allocated for th... | LLM in a flash |
Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste
Lespiau, Maria Tsimpoukelli, Nikolai Grigorev,
Doug Fritz, Thibault Sottiaux, Mantas Pajarskas,
Toby Pohlen, Zhitao Gong, Daniel Toyama, Cy-
prien de Masson d’Autume, Yujia Li, Tayfun Terzi,
Vladimir Mikulik, Igor Babuschkin, Aidan Clark,
Diego de Las Casas, Aurelia Guy,... | LLaMA- Open and Efficient Foundation Language Models |
sentence?. Here we point out the target domain by replacing {DOMAIN} with domain names such
as biomedicine, finance, or law. Besides, we task the language model with defining concepts
using the mining pattern and input-output template in Table 2.
Natural Language Inference concerns how two sentences relate, typically a... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
to online harassment, and 66 percent has witnessed such behaviors directed at
others. (These numbers are higher than the percentage of people who use digital
media to actively comment on or share news online.) Almost one in five
(18 percent) reported having been subjected to particularly severe forms of
online harassmen... | Social_Media_and_Democracy |
[216] Xinnuo Xu, Ondřej Dušek, Verena Rieser, and Ioannis Konstas. 2021. AGGGEN: Ordering and Aggregating while
Generating. roceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL2021) (2021).
[217] Yan Xu, Etsuko Ishii, Samuel Cahyawijaya, Zihan Liu, Genta Indra Winata, Andrea Madot... | SurveyofHallucinationinNatural Language Generation |
Trends to Watch: Creator Economy
22
NFT creator royalties paid by marketplace (on Ethereum)
NFT creators
have earned
more than
$1.9 billion
in royalty
revenues
Transfer-based royalties are
under fire due to a lack of
viable on-chain enforcement.
The industry is exploring
alternative solutions.
$300M
... | State-of-Crypto2023 |
settings, exemplars are selected randomly from the dataset excluding the example being evaluated.
We find that Flan-PaLM generates responses prefixed with articles (e.g. "the engineer"), particularly in
zero-shot settings. We find this in Flan-PaLM in 78% to 86% of responses and in 96% of Flan-T5-XXL responses.
After adap... | Scaling Instruction-Finetuned Language Models |
In order to evaluate the run time efficiency of FORGE, we chose to focus on the smallest dataset of the benchmark study
in Sect. 5.2, namely adult. We (A) drew stratified subsamples and (B) drew covariate subsets. For step (B), the target
variable is always included. We select an equal number of continuous/categorical co... | Adversarial Random Forests for Density Estimation and Generative Modeling |
[24] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter
Abbeel, and Ben Poole. Zero-shot text-guided object genera-
tion with dream fields. In Proceedings of the IEEE/CVF Con-
ference on Computer Vision and Pattern Recognition, pages
867–876, 2022. 2
[25] Heewoo Jun and Alex Nichol.
ing conditional 3d implicit func... | Wonder3D |
models. arXiv preprint arXiv:2106.04426, 2021.
Sascha Rothe, Jonathan Mallinson, Eric Malmi, Sebastian Krause, and Aliaksei Severyn. A simple
recipe for multilingual grammatical error correction. arXiv preprint arXiv:2106.03830, 2021.
Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande:... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Representation of global body orientation: Global
subject orientation can be represented as body rotation or,
equivalently, camera rotation. We choose to rotate the cam-
era in order to keep the estimated bounding box when sub-
ject orientation changes. Specifically, we uses axis-aligned
bounding boxes because for ease... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
Zi→ Xi). The HCLT structure is then compiled into a PC
that encodes the same probability distribution. We used
the hybrid mini-batch + full-batch EM as described in Sec. 4.1. For all experiments, we trained the
PCs with 100 mini-batch EM epochs and 100 full-batch EM epochs. Please refer to Appendix B.4
for hyperparamet... | Tractable Regularization of Probabilistic Circuits |
models have shown the emergent ability of planning in forms of goal reformulation [99; 124], task
decomposition [98; 125], and adjusting plans in response to environmental changes [100; 126]. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
B
a
r
d
i
s
b
a
s
e
d
o
n
a
l
i
g
h
t
w
e
i
g
h
t
a
n
d
o
p
t
i
m
i
z
e
d
v
e
r
s
i
o
n
o
f
(
s
h
o
rt
f
o
r
"
L
a
n
g
u
a
g
e
M
o
d
e
l
s
f
o
r
D
i
a
l
o
g
u
e
A
p
p
l
i
c
a
t
i
o
n
s
"
)
,
a
n
d
,
s
i
m
i
l
a
r
t
o
m
o
s
t
L
L
M
s
t
o
d
a
y
,
w
a
s
p
r
e
-
t
r
a
i
n
e
d... | An overview of Bard- an early experiment with generative AI |
Large language models are not fair evaluators. arXiv preprint arXiv:2305.17926 (2023).
[198] Rose E Wang and Dorottya Demszky. 2023. Is ChatGPT a Good Teacher Coach? Measuring Zero-Shot Performance
For Scoring and Providing Actionable Insights on Classroom Instruction. arXiv preprint arXiv:2306.03090 (2023).
[199] X... | ASurveyonEvaluationofLargeLanguageModels |
hardware and accelerators such as GPUs.
For this particular checkpoint, note that the mode tags we used are [NLG] (X-denoiser), [NLU] (R-denoiser)
and [S2S] (S-denoiser). So add that at the start of the inputs of your examples. | UL2- Unifying Language Learning Paradigms |
3.3 Action
Action
Textual Output §3.3.1
Learning tools
Toolformer [92], TALM [326], Instruct-
GPT [24], Clarebout et al. [327], etc.
Tools §3.3.2
Using tools
WebGPT [90], OpenAGI [211], Visual
ChatGPT [328], SayCan [179], etc.
Making tools
LATM [329], CREATOR [330],
SELF-DEBUGGING [331], etc.
LLM-based
Embod... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
generally, via a laptop, tablet or smartphone. These, assistive devices are not always quick to access
and often carry with them a stigma [2].
Wearables, such as smartwatches and smart glasses, have the potential to provide a range of sensors
and modes of input/output within an unobtrusive, commonplace form factor... | informatics-phd-projects-2022-23 |
• Accelerating research and innovation: The open-source
model fuels progress in research and development within
the AI domain. It allows researchers to leverage existing
models, thus nurturing a faster progression of innovation
and scientific discovery.
• Enhancing education: Open-source FinLLMs serve as ro-
bust educ... | FinGPT-Open-SourceFinancialLargeLanguageModels |
5.4 Speaker Diarization
5.4.1 Task Description
Speaker diarization is a critical component in the analysis of multi-speaker audio data, and it
addresses the question of "who spoke when." The term "diarize" refers to the process of making a
note or keeping a record of events, as per the English dictionary. A traditional... | AReviewofDeepLearningTechniquesforSpeechProcessing |
3.2 Problem formulation
Given a dataset of transcribed speech (x, y) where x and y denote an audio sample and its transcript,
respectively, the goal is to build a single model that can perform many text-guided speech generation
tasks through in-context learning. We propose to train such a generative model on the text-... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
Programme.
assessments.
2.8.3
Initial Entry
1. RPL may be considered for initial entry to a UCL taught or research Programme where a
student does not meet the standard entry requirements as defined in Section 2: Entrance
Requirements and Chapter 5: Research Degrees Framework e.g. a student holds an
interna... | UCL Academic Manual |
Q: A car is running at a speed of 96kmph. What distance will it cover in 14sec? Options: A:378m B:350m C:380m
D:200m E:250m
A: Reasoning process: 1. We are given that the car is running at a speed of 96 km per hour. 2. We want to find the
distance it will cover in 14 seconds. 3. We need to convert both km and hours to m... | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
.
.
.
.
.
.
.
.
.
F. The Details of other Creative Tasks
F.1. The Details of Cloud Guessing Game (CGG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
F.2. The Details of Divergent Association Task (DAT) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
G. The Analysis for Self-R... | Let’sThinkOutsidetheBox |
128:16
• Villa et al.
attributes on the perception of augmented humans and technological attributes. In Survey #3, we could also
show concurrent validity in that attitudes toward augmented humans can predict whether participants are
willing to use augmentation technologies themselves. Therefore, positive attitudes re... | Society’sAttitudesTowardsHumanAugmentation |
pose to regularize via prompt consistency, encour-
aging consistent predictions over the prompts. This
allows either finetuning the model with extra un-
labeled training data, or direct application at infer-
ence time. While SELF-INSTRUCT has some simi-
larities with the self-training literature, most self-
training met... | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
Table 14: HumanEval single line infilling. pass@1 on the infilling benchmarks from Fried et al. (2023)
and Bavarian et al. (2022). Evaluated with greedy decoding in both prefix-suffix-middle (PSM) and suffix-
prefix-middle (SPM) format. LCFT refers to long-context fine-tuned models. Numbers are reported from
Bavarian e... | CodeLlama2 |
Question: Show the status of the city that has hosted the greatest number of
competitions.
20
SQL: SELECT city.status FROM city JOIN farm_competition ON city.city_id =
farm_competition.host_city_id GROUP BY farm_competition.host_city_id ORDER
BY COUNT(*) DESC LIMIT 1
CREATE TABLE customers (
customer_id number ,
cu... | Teaching Large Language Models to Self-Debug |
If you are a founder and would like to meet, shoot us a note at sonya@sequoiacap.com and grady@sequoiacap.com.
We can’t wait to hear your story.
PS: This piece was co-written with GPT-3. GPT-3 did not spit out the entire article, but it was responsible for combating writer’s block, generating entire sentences and parag... | Generative AI A Creative New World Sequoia Capital |
Admissions in Student & Registry Services.
3. Departments offering applicants fully-funded places at UCL are responsible for ensuring that
funding is available for the duration of the programme.
3.9.3 Conditional Offers
1. Conditional offers based on future examination performance may be issued by UCL. In s... | UCL Academic Manual |
models to do a single simplification step of turning an N digit problem into an N − 1 digit problem
and a partial solution, similar to the method of teaching addition to schoolchildren. In this setup,
performing “slow” addition refers do performing a single step of simpliciation, not fully solving the
problem. Figure 2... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
Law (NetzDG)
combating, 71–75, 77
defining, 56–59, 75
detecting, 59–61, 76
introduction, 56
offline consequences of, 68–71, 77
prevalence of, 66–68
producers of, 61–64, 76
targets of, 64–66, 68–69, 76
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
338
Index
Hate Speech Code of ... | Social_Media_and_Democracy |
Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima
https://yoheinakajima.com/task-driven-autonomous-agent-utilizing-gpt-4-pinecone-and-langchain-for-diverse-applications/
6/8 | Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications – Yohei Nakajima |
ZENY: That is correct, Socart.
SOCART:
. . . but do we really know why there are so
few NLP papers about negative results? See, in
NLP, negative results are much harder to estab-
lish than positives. If I want to show that self-
attention or weight averaging does not lead to
improvements for some problem, I need to sh... | A Two-Sided Discussion of Preregistration of NLP Research |
Remuneration is based on the collective agreement for the public service of the federal states (TV-L). At the end of the year, the TV-L provides for a special annual payment. In addi‐
tion, we offer a separate supplementary pension in the form of a company pension (VBL).
Equality with its facets of equal opportunities... | _2 Doctoral Researcher (m_w_d) in the field of Large Language Models (LLM) for Software Engineering_ - Technische Universität Clausthal - DAAD |
9
4.1 LLM Personality Characterization
The methodology for characterizing LLM personality and quantifying its ability to
coherently emulate human personality traits consists of two steps. First, we administer
psychometric tests to LLMs and collect the scores. Second, those scores are used to
establish construct vali... | PersonalityTraitsinLargeLanguageModels |
ward and convolutional networks, stacked on pre-trained
text embeddings modules publicly available via TensorFlow
Hub5. The embeddings coming from the TensorFlow Hub
modules may be frozen or fine-tuned. The full search space
is described in the appendix. For each task, we run AutoML
for one week on CPUs, using 30 machin... | Parameter-Efficient Transfer Learning for NLP |
1Note that not all capabilities of PaLM 2 are currently exposed via PaLM 2 APIs.
3
Figure 1: Performance of PaLM 2 and PaLM on the latest available professional language proficiency exams. We
used exams for each language to test a C2 (mastery or advanced professional) level proficiency following the CEFR
definition. We... | PaLM 2 Technical Report |
𝑝 > 0 else 0
end for
Compute 𝑛@𝑘 for this problem as the average of all solved𝑝.
𝑝 ∼Hypergeometric(𝑒𝑝, 𝐾 − 𝑒𝑝, 𝑘) ⊲ # samples out of 𝑘 which pass examples tests.
⊲ Only submit samples that pass the example tests.
𝑝 ∼Hypergeometric(𝑠𝑝, 𝑒𝑝 − 𝑠𝑝, 𝑛(cid:48))
⊲ # correct solutions out of 𝑛(cid:48) sub... | alphacode |
1
Introduction | Translatotron3 |
40
method for training foundation models for RL, R3M [Nair et al., 2022], combines time-
contrastive and video-language alignment objectives. VIP and R3M are trained on the
large Ego4D dataset [Grauman et al., 2022], while MVP combines Imagenet, Ego4D, and
additional hand manipulation data. Majumdar et al. propose VC... | A Cookbook of Self-Supervised Learning |
Describe this image.The image shows a cat wearing a bluecookie monster costume sitting on acouch with a plate of cookies infront of it. The cat's mouth is openand it appears to be enjoying thecookies. The caption reads, `anotherwild saturday night.`The image is a cute and funnydepiction of a cat in a cookiemonster cost... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
Wickham-Jones, M. (2018). This 1950 political science report keeps popping up in the
July 24. www
news: Here’s
.washingtonpost.com/news/monkey-cage/wp/2018/07/24/this-1950-political-
science-report-keeps-popping-up-in-the-news-heres-the-story-behind-it/
story behind it. Washington Post,
the
Zaller, J. R. (1992). Th... | Social_Media_and_Democracy |
19
(a) 280M
(b) 1B
20
Figure 8: Exponential moving average of domain weights throughout a DoReMi run for 280M and
1B reference/proxy models. In the beginning of the run, the domain weights change quickly and
then become more stable after 50k steps. This suggests that 1) smaller compute budgets may require
drastica... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
[26] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng,
Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas
Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter.
Inner monologue:
Embodied reasoning through planning with language models. a... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
Figure 31
harmless data. (right) Learning curves on the harmlessness test set.
(left) Learning curves on the helpfulness test set when training on a mix of static helpful and
In all phases, we only train over one iteration to mitigate overfitting.
A.3 Scaling of PM with Model and Dataset Size
A major question is how... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
However for ambiguous contexts, PaLM 2 was 0.6% accurate overall, and generated a response that named either the
bias target or non-target even though this information could not be accurately determined from the prompt. Further,
when doing this, it was more likely to produce a response that reinforced a social bias. Th... | PaLM 2 Technical Report |
scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1
[40] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing
scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 4, 7
[41] Michael Niemeyer, Lars Mescheder, Michae... | I M Avatar- Implicit Morphable Head Avatars from Videos |
8.5 SPECULATIVE DECODING
Table 9 reports the relative latency of the medium.en and large-v2 models with speculative decoding.
We compare the latency using either the smallest Whisper checkpoints or the Distil-Whisper models
as the assistant. Since the outputs of the original Whisper models are obtained exactly, we rep... | DISTIL-WHISPER |
Active adversaries are currently exploiting this flaws to evade the detection mechanisms placed in
current OSN. This applies to several application domains such as: i) the use of spear phishing or
malware attacks to enable cyber-dependent crime, ii) the use of coordinated harassment campaign to
deliver harmful or de... | informatics-phd-projects-2022-23 |
3 Experimental Setup
Models We evaluate three PLMs: mBERT (De-
vlin et al., 2019), XLM-RoBERTa/XLM-R (Con-
neau et al., 2020), and mT5 (Xue et al., 2021).7
All three models were trained with a masked lan-
guage modelling objective. mBERT differs from
XLM-R and mT5 in including a next sentence pre-
diction objective (De... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
4
1.3
1.8
2.2
1.0
1.3
1.3
16
1.5
1.2
1.3
1.0
0.8
0.8
769
808
856
1550
1589
1669
experiments. Therefore, our final KD training objective is a weighted sum of the PL and KL terms
only.
D.3 BATCHED SPECULATIVE DECODING | DISTIL-WHISPER |
or even exceeds human performance on multiple tasks. This evaluation requires human evaluators
to actually test and compare the performance of the models, not just evaluate the models through
automated evaluation metrics. Note that even human evaluations can have high variance and
instability, which could be due to cul... | ASurveyonEvaluationofLargeLanguageModels |
musical structure, thus enhancing its quality.
A slightly different task, relates to reconstructing music from instrument per-
formances that lack the accompanying audio (Gan et al., 2020; Koepke et al.,
2020; Su et al., 2020a,b), such as generating piano music from a video of fin-
ger movements on the piano. Recen... | Video2Music |
facilitate connections with readers to both spread and retrieve news (Gonzales
and González 2017). Gonzalez and Gonzalez, exploring the case of a service
known PolitiBot, which operated during the 2016 Spanish election over
Telegram, write that the true journalistic potential of the program was to
share relevant news (... | Social_Media_and_Democracy |
Within the scope of AI, the Turing Test [182], a widely recognized test for assessing intelligence
by discerning if responses are of human or machine origin, has been a longstanding objective in AI
evolution. It is generally believed among researchers that a computing machine that successfully
passes the Turing Test ca... | ASurveyonEvaluationofLargeLanguageModels |
4.3 Semi-supervised Learning
Semi-supervised learning can be viewed as a process of optimizing a model using both labeled and
unlabeled data. The set of labeled data points, denoted by 𝑋𝐿, contains 𝑁𝐿 items, where each item
is represented as (𝑥𝑖, 𝑦𝑖) with 𝑦𝑖 being the label of 𝑥𝑖. On the other hand, the set... | AReviewofDeepLearningTechniquesforSpeechProcessing |
appropriate guardrails, open LLM development poses a higher risk of misuse, as increased model
access also increases the likelihood of harm caused by the model. Even though a released API can be
shut down, once the model weights are released, it is nearly impossible to retract them. Therefore, it
is crucial to discuss ... | StarCoder_paper (1) |
Table 7. The full configuration details for Pythia training. Exact model config files are also made available via our Github repository.
Configuration values marked with “–” differ between models. Table 1 provides particular model dimensions. Additionally,
some modifications are necessary to enable appropriate parallelism:... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Router Z-Loss
Fraction Stable
4/6
3/3
3/3
Quality (↑)
-1.755 ±0.02
-4.206 ±0.17
-1.741 ±0.02
Table 4: Constraining weight updates and router logits. Constraining the update clipping in
Adafactor improves stability, but at a catastrophic loss of quality. Looser clipping values did not
reliably stabilize training so ... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
This literature review details empirical work on how the bot as an internet-
based tool, and computational propaganda as a political communication
strategy, function in relationship to social media and democracy. As Luceri
et al. (2019) argue, “the presence of social bots does not show any sign of
decline despite the a... | Social_Media_and_Democracy |
eter δ = 0 and use the default mtry = (cid:98)√
5.1 Simulation
FORGE recreates visual patterns. We begin with a sim-
ple proof of concept experiment, illustrating our method on
a handful of low-dimensional datasets that allow for easy
visual assessment. The cassini, shapes, smiley, and
twomoons problems are all three... | Adversarial Random Forests for Density Estimation and Generative Modeling |
02468101-Shot GEM XSum, SGD, TOT Avg. Rouge-L6062646668707274Finetuned SuperGLUE ScoreGPT-like(Dec)PrefixLM (Dec)SpanCorrupt (Dec)UL2 (Dec)PrefixLM (EncDec)T5 (EncDec)UniLM (EncDec)UL2 (EncDec) new ones; namely X-denoising (extreme denoising) which considers extreme span lengths and corruption
rates, S-denoising (seque... | UL2- Unifying Language Learning Paradigms |
Principal-agent VCG contracts - ScienceDirect
Journal of Economic Theory, Volume 204, 2022, Article 105498
Show abstract
Research article
Research article
Journal of Economic Theory, Volume 202, 2022, Article 105458
Show abstract
Journal of Economic Theory, Volume 201, 2022, Article 105418
Show abstract
☆
A one-... | Principal-agent VCG contracts - ScienceDirect |
Researchers are also paying attention to data-centric knowledge injection [113]
when adapting a general-purpose foundation model towards a specific domain such
19
as healthcare. The knowledge injection can be achieved by fine-tuning on domain-
specific textbooks, publications, and instructions. As an identified drawback... | Beyond Efficiency |
pre(g(a)) ⊆ s2 there must be some s1 ∈ S1 such that pre(a) ⊆ s1 and s1[V C] = s2, i.e. s1 ∈ f (s2). Then (cid:3)s1, t1, a(cid:4) ∈ E1, where | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Table 15: Results on MMLU and BBH using FLAN-UL2. † denotes checkpoint that we release.
Table 15 reports the results on MMLU and BBH (Suzgun et al., 2022). Generally, the performance of FLAN-
UL2 20B is pretty competitive and outperforms Flan-T5 XXL by +1.8% on the test set and +4.7% on MMLU
dev. The Big-Bench hard sc... | UL2- Unifying Language Learning Paradigms |
In the context of TMP, tensors can be split along their rows or columns,
enabling concurrent execution of matrix multiplication operations across all split parts.
Megatron-lm [16] employs parallelization techniques for matrix multiplication oper-
ations within both the multi-layer perceptron (MLP) and the self-attentio... | Beyond Efficiency |
We now discuss the problem of RLHF robustness. A fully robust PM would agree with humans on distribu-
tions of dialogues quite different from those encountered during PM training (i.e. different from those created
by crowdworker interactions with our deployed AI assistants). However, we do not expect that our PMs are
s... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
2.2 Architecture
We adopt a similar model architecture to Llama 2 (Touvron et al., 2023b). We use a Transformer
architecture based on Vaswani et al. (2017) with the following details:
2
Table 1: The details of model architecture
Hidden size
2,048
Intermediate Hidden Size Context Len Heads Layers Vocab size
5,632... | TinyLlama |
Now, grade the student’s completion in terms of grammar, creativity, consistency with the story’s beginning and
whether the plot makes sense. Moreover, please provide your best guess of what the age of the student might be,
as reflected from the completion. Choose from possible age groups: A: 3 or under. B: 4-5. C: 6-7... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
[37] Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDermott, Manel
Zarrouk, and Alexandra Balahur. Www’18 open challenge: financial opinion mining and
question answering. In Companion proceedings of the the web conference 2018, pages 1941–
1942, 2018.
[38] Tomas Mikolov, Kai Chen, Gregory S. Corra... | E5 |
also include full interactivity, like the ability to see who
shared and commented, as well as full functionality, with
other ways to respond as a participant. Other future stud-
ies might look at the source of the flags, whether govern-
ment, the social media company, or other users, to see which
source is most effe... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
We design new mugs shown in theimage. Can you help us write anadvertisement?Sure, here's an advertisement for themugs:Introducing our new line of mugsfeaturing adorable cat designs! Thesemugs are perfect for coffee loversand cat enthusiasts alike. Made fromhigh-quality ceramic, each mug isdurable and dishwasher safe. T... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
Autonomy. Autonomy means that an agent operates without direct intervention from humans
or others and possesses a degree of control over its actions and internal states [4; 113]. This
implies that an agent should not only possess the capability to follow explicit human instructions for
task completion but also exhibit ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Graph Representation of Speech Data. The first step in using GNNs for speech processing
is representing the speech data as a graph. One way to do this is to represent the speech signal as a
sequence of frames, each representing a short audio signal segment. We can then represent each
frame as a node in the graph, with ... | AReviewofDeepLearningTechniquesforSpeechProcessing |
0.466
0.480
-
SPICE
0.144
0.176
-
0.177
0.182
Model
ORG-TRL [58]
MV-GPT [43]
GIT [48]
mPLUG-2 [52]
CoDi (Ours)
B@4 METEOR CIDEr
50.9
43.6
60.0
48.9
75.9
54.8
57.8
80.3
74.4
52.1
28.8
38.7
33.1
34.9
32.5 | Any-to-Any Generation via Composable Diffusion |
4
Universal Self-Consistency for Large Language Model Generation
4.2 MAIN RESULTS
Mathematical reasoning. For mathematical reasoning benchmarks, we compare USC against
the standard self-consistency in Table 1. For the standard self-consistency, we employ a regular
expression matching to extract the final answer on ... | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
Figure 2: DoReMi optimizes domain weights with a small model (280M params) and uses these
domain weights to train a much larger model (8B params, 30x larger). Here, optimizing the do-
main weights (training a small model twice) takes 8% of the compute of training the large model.
DoReMi improves average one-shot downst... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
Kreiss, D., & McGregor, S. C. (2019). The “arbiters of what our voters see”: Facebook
and Google’s struggle with policy, process, and enforcement around political
advertising. Political Communication, 36(4), 499–522. https://doi.org/10.1080
/10584609.2019.1619639
Libert, T., Graves, L., & Nielsen, R. K. (2018). Change... | Social_Media_and_Democracy |
show that our proposed LaMini-LM are on par
with competitive baselines while being nearly
×10 smaller in size.1 | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
Scaling laws generally only predict a model’s pretraining
test loss, which measures the model’s ability to correctly
predict how an incomplete piece of text will be continued.1
While this measure is correlated with how useful a model
will be on average across many practical tasks (Radford
et al., 2019), it is largely n... | Eight Things to Know about Large Language Models |
The starting salary for the position is 2 693, 21 EUR per month. In
addition to the salary, the contract includes occupational health benefits,
and Finland has a comprehensive social security system. The annual total
workload of research and teaching staff at Aalto University is 1 612 hours.
The position is located at ... | Doctoral researcher position in Human-Computer Interaction _ Human-AI Interaction _ Aalto University |
dimensionality grows linearly with the number of input symbols.
It has become the de facto standard in NLP to define a fixed width embedding matrix E(cid:96) ∈ R|Σ|×d
(cid:96) b.6 Rather than using a dot-
where width d is a hyperparameter, and embed each symbol as ET
product, this process is implemented as a lookup opera... | MULTI HASH EMBEDDINGS IN SPACY |
3.4 Harmfulness Scores on Held Out Prompts
For our internal evaluation of Claude models, we gauge harmfulness using a held-out set of 328 prompts that
include representative examples from our red-teaming work [4] and various AI model ‘jailbreaks’ that have
been discussed online. We then compare HHH preference model sc... | ClaudeModels |
Examiners, Australian Journal of Educational & Developmental Psychology, v4 p126-145
Kothari, C.R. (2006). Research methodology: Methods & techniques. India: New Age International Publishers
Krathwohl, David R. 1988. How to Prepare a Research Proposal: Guidelines for Funding and Dissertations
in
the Social... | How to Write Your PhD Proposal- A Step-By-Step Guide |
engagement, even if most people never encounter it. Similarly, in an analysis of
junk news during the EU parliamentary elections in 2019, Marchal et al. (n.d.)
show that less than 4 percent of EU-related news links circulating on Twitter
during this time were from disreputable sources. Notably, though, the Polish
Twitt... | Social_Media_and_Democracy |
monly used key finding algorithms: Krumhansl-Schmuckler (Krumhansl, 2001),
Temperley-Kostka-Payne (Temperley, 2007), and Bellman-Budge (Bellmann,
2006). To consolidate they key detection results, we implemented a voting
method that considers the output from all three algorithms. This ensemble
approach enhances the ... | Video2Music |
5
Interpretability
Understanding the inner workings of deep neural networks and language models in particular is a major challenge
in this field of study. For example, it is often difficult to assign a specific function to a given component of a
neural network. This may be because, contrary to our intuition based on ... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
ballot for the Republican challenger for U.S. Senate in Georgia,” “Smith for Congress,” “Bill
McKay in ’94.”
7 Prior to Citizens United v. Federal Election Commission 558 U.S. 310 (2010), express advocacy
messages as outlined in this section (on television, in print, on radio, online, and so on, regardless
of format) ... | Social_Media_and_Democracy |
4.6 Hours on social media and news media
A Pearson correlation identified a trend for participants
in how their rating changed after viewing the Tweets with
a flag for a suspected bot account and when there was a
Fig. 5 Differences in news media consumption on opinions about COVID-19 death count
1 3 32 Page ... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
o
f
t
h
e
e
a
r
l
y
,
p
r
o
m
i
s
i
n
g
a
p
p
l
i
c
a
t
i
o
n
s
o
f
B
a
r
d
a
b
o
v
e
,
w
h
i
c
h
r
e
q
u
i
r
e
a
s
u
s
t
a
i
n
a
b
l
e
w
e
b
c
o
n
t
e
n
t
e
c
o
s
y
s
t
e
m
.
W
e
’
r
e
c
o
m
m
i
tt
e
d
t
o
i
n
n
o
v
a
t
i
n
g
i
n
t
h
i
s
s
p
a
c
e
r
e
s
p
o
n
s
i
b
l
y
,
i... | An overview of Bard- an early experiment with generative AI |
Lgeo = || log(R∗RT )||, R = PoseNet(ψrnd),
(21)
where we find learning to predict rotation is sufficient for
initializing the root body pose.
In practice, we set the
initial object-to-camera translation to be a constant T =
(0, 0, 3)T . We run pose CNN on each test video frame to
obtain the initial root poses Gt
delt... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
CREATE TABLE customers (
customer_id number ,
customer_first_name text ,
customer_last_name text ,
customer_address text ,
customer_phone text ,
43
customer_email text ,
other_customer_details text ,
primary key ( customer_id )
)
insert into customers (customer_id, customer_first_name, customer_last_name,
customer_a... | Teaching Large Language Models to Self-Debug |
Domain 2: Consumer Confidence
The consumer confidence setting is helpful in answering RQ3 (which types of opinions are most affected by media consumption).
Across all questions, there is only a weak correlation (r=0.104, CI(0.066, 0.142)), but categorizing the questions by question
type is illuminating. The consumer confi... | Language models trained on media diets can predict public opinion |
A data-efficient pretraining strategy for the PM-VLN
module consists of pretraining submodules of the PM-VLN
on auxiliary tasks (ϕ1, ϕ2). We denote the two datasets as
(Dϕ1,Dϕ2 ) and a training partition as DT rain. In ϕ1, the
gP M T P submodule is pretrained on TR-NY-PIT-central - a
new set of path traces. Path traces... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.