text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How does llm safety training
fail? arXiv preprint arXiv:2307.02483, 2023.
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny
Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advan... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
11We do not know whether 1
2 is tight.
10
In this example, Welai(v) = q + i − (i − 1)(γ + (cid:15)). Thus, the socially efficient action is aq,
while the action that minimizes welfare is a1.
In Appendix A.2 we show that, when using a
first price contract, the agent takes a1 in all equilibria. The intuitive reason for ... | Incomplete Information VCG Contracts for Common Agency |
Causal judgement
n/a
English Proverbs
n/a
Implicatures
n/a
Nonsense words
grammar
Rhyming
Which word in the following sentence is a verb?
The grilshaws bolheavened whincely.
Reason: Linguistically-typical suffixes (i.e. -ed
for a verb).
What rhymes with ’cruise’?
Reason: Model cannot rely on spelling or au-
dio... | AreEmergentAbilitiesinLarge Language Models just In-Context |
Hwang, Pearce, and Nanis (2012) explore the ways that bots can be used
as a social prosthesis or scaffolding for connecting networks of people that
might not otherwise communicate. They argue, citing natural bot-driven
experiments on social media, that bots can be effectively used to parse
information on a social netwo... | Social_Media_and_Democracy |
How sensitive are sparse and dense models to the fine-tuning protocol? We study two hyperparam-
eters: the batch size and the learning rate. We pretrain a Dense-L and ST-MoE-L on 500B tokens
of C4 and then fine-tune on SuperGLUE. Figure 6 summarizes our experiments with the full data
presented in Table 20 (Appendix F). A... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
3. Method
In this section, we first
Given a large 2D image collection, our goal is to learn a
generative model of diverse 3D human avatars with realistic
appearance and geometry, while enabling control over pose
and identity. An overview of our method is shown in Fig. 2.
introduce an efficient and
articulation-aware 3D... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
parameters:
6Below, there will be many other examples of completions of stories from outside of the training set.
7
Summary: A cat performs a new trick for her friends but starts shivering. Her friends give her a warm hug and she realizes their love is the
Sentence: The cat started to do her trick, but then someth... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
SpaceCamera SpaceAn image at time txt<latexit sha1_base64="fglfFNNfFJ1LSzytA6p8EIsF9U4=">AAAB9XicbVDLSsNAFL3xWeur6tLNYBFclUQEXRbcuKxgH9KmZTKdtEMnD2Zu1BLyH25cKOLWf3Hn3zhps9DWAwOHc+7lnjleLIVG2/62VlbX1jc2S1vl7Z3dvf3KwWFLR4livMkiGamORzWXIuRNFCh5J1acBp7kbW9ynfvtB660iMI7nMbcDegoFL5gFI3U7wUUx56fPmX9FLNBpWrX7BnIMnEKUoUCjUHlqze... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
3.6 Massive Multitask Language
Understanding
The massive multitask language understanding
benchmark, or MMLU, introduced by Hendrycks
et al. (2020) consists of multiple choice questions
covering various domains of knowledge, includ-
ing humanities, STEM and social sciences. We
evaluate our models in the 5-shot settin... | LLaMA- Open and Efficient Foundation Language Models |
Rebuffi, S.-A., Bilen, H., and Vedaldi, A. Learning multiple
visual domains with residual adapters. In NIPS. 2017.
Rosenfeld, A. and Tsotsos, J. K.
Incremental learning
through deep adaptation. IEEE transactions on pattern
analysis and machine intelligence, 2018.
Rusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer... | Parameter-Efficient Transfer Learning for NLP |
Third, CDA 230 does not preclude more dramatic interventions that would
change the actual flow of information through platforms. As a means of limiting
the influence of online platforms in shaping public discourse, policymakers have
called for a form of “net neutrality” to apply to the content layer of the Web, such
that... | Social_Media_and_Democracy |
30
Figure 24 This figure shows individually-normalized histograms of the distribution of PM scores that our
online HH PM assigns to samples written by professional writers, alongside samples from our HH and
helpfulness-only online RLHF models. Our PM prefers our models’ samples to those written by the hu-
man writers,... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
2https://cloud.google.com/
natural-language/docs/basics#entity_
analysis
4939Transformer LayersEntity Memory...Transformer LayersQ1035: Charles Darwin Charles [MASK] [MASK] published the Origin of the Species in 1859 .BIBIII ei (cid:54)= e∅, the pseudo embedding hmi (Equatio... | Entities as Experts- Sparse Memory Access with Entity Supervision |
and HuggingGPT (Shen et al., 2023) extend APIs and tasks to broader scenarios, including multimodal
models for visual tasks, local software, and cloud service APIs. OpenAI also proposed its official tool
library, ChatGPT Plugins 23, to empower ChatGPT with other applications. By simply providing APIs with
descriptions, ... | Tool Learning with Foundation Models |
cant improvements over state-of-the-art methods on dynamic
scene datasets, and also apply our approach to in-the-wild
videos with challenging camera and object motion, where
prior methods fail to produce high-quality renderings.
1. Introduction
Computer vision methods can now produce free-
viewpoint renderings of sta... | DynIBaR-NeuralDynamicImage-BasedRendering |
madaka, Jianyu Huang, Hector Yuen, et al. 2019. A study of BFLOAT16 for deep learning training. arXiv preprint arXiv:1905.12322 (2019).
[124] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei.
2020. Scaling laws for neural lang... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Table 1. Attitudes towards COVID-19: feature importances in regressions of media diet scores against survey response proportions. For one
media diet subpopulation, the “attention to news” feature is the percentage of respondents who answered that they were paying “very close”
attention to coronavirus-related news. The ... | Language models trained on media diets can predict public opinion |
V
o
x
P
o
p
u
l
i
.
e
n
10.9
11.6
9.3
9.5
8.1
8.2
7.8
7.6
7.2
7.0
F
l
e
u
r
s
.
e
n
u
s
10.0
12.4
7.1
8.9
5.3
6.1
4.5
4.4
4.5
4.2
Table 9. English transcription WER (%) with beam search and temperature fallback
Robust Speech Recognition via Large-Scale Weak Supervision
23
D.2. Multilingual Transcription
D.2.... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
Shape Modeling Procedure. We recruit six professional
artists to create 3D corresponding character models using
Blender according to the collected reference images. The key
to building a linear parametric shape model lies in maintain-
ing a unified mesh topology. To achieve this, all six artists
are required to craft 3... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
Worldwide (WW) net sales
WW net sales -- Y/Y growth, excluding F/X
WW net sales -- TTM
WW net sales -- TTM Y/Y growth, excluding F/X
Operating income
F/X impact -- favorable
Operating income -- Y/Y growth (decline), excluding F/X
Operating margin -- % of WW net sales
Operating income -- TTM
Operating income -- TTM Y/Y ... | AMZN-Q3-2023-Earnings-Release |
scripts or symbolic music representations.
MusicLM builds on top of AudioLM with three important
additional contributions: (1) we condition the generation
process on a descriptive text, (2) we show that the condition-
ing can be extended to other signals such as melody, and
(3) we model a large variety of long music se... | MusicLM |
4.1.2. Transformer Encoder
In the process of encoding input vectors that represent videos, denoted
as Inputvideo, a Transformer Encoder with L layers is utilized. Each layer
22
l within the range 1 ≤ l ≤ L takes the current contextual representation
H (l−1) and transforms it into the subsequent output H (l) throug... | Video2Music |
on low resource language summarization and
develops a novel metric, mFACT to evaluate the
faithfulness of non-English summaries, leveraging
translation-based transfer from multiple English
faithfulness metrics.
It is developed from four
English faithfulness metrics. They study hallucina-
tion in a cross-lingual transfe... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
and 3D cartoon animation. Experimental results demonstrate
the practicality of 3DBiCar and RaBit.
Parametric Texture Modeling. Traditionally, textures are
modeled as a linear subspace using the similar idea of body
blendshape models. Blanz et al. [5] represent the face ap-
pearance in per-vertex colors and parameterize... | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
In the summer of 2018, Facebook initiated a data-sharing initiative with
Social Science One, an academic effort seeking to make Facebook and other
industry data available to the larger scientific community (King and Persily
2019).3 Through the Commission created by Social Science One, which was
funded not by Facebook bu... | Social_Media_and_Democracy |
u
s
e
c
a
s
e
s
,
t
h
e
o
n
l
y
f
e
a
t
u
r
e
t
h
a
t
m
a
t
t
e
r
e
d
w
a
s
r
e
a
l
-
t
i
m
e
c
o
l
l
a
b
o
r
a
t
i
o
n
,
w
h
i
c
h
a
l
l
o
w
e
d
G
S
h
e
e
t
s
t
o
c
a
p
t
u
r
e
m
a
r
k
e
t
s
h
a
r
e
d
e
s
p
i
t
e
h
a
v
i
n
g
f
e
w
e
r
f
e
a
t
u
r
e
s
t
h
a
n
M
i
c
r
o
s
o
f
t
... | Product-Led AI _ Greylock |
Timo Schick, Sahana Udupa, and Hinrich Schütze. Self-diagnosis and self-debiasing: A proposal for reducing
corpus-based bias in NLP. CoRR, abs/2103.00453, 2021. URL https://arxiv.org/abs/2103.00453.
Thomas Scialom, Tuhin Chakrabarty, and Smaranda Muresan. Continual-T0: Progressively instructing
50+ tasks to language ... | Scaling Instruction-Finetuned Language Models |
4.4 Metrics
We evaluate Bash generation using the tem-
plate match metric—which performs some basic
normalization—provided with the dataset. We eval-
uate Python using CodeBERTScore (Zhou et al.,
2023), which has been shown to be a high quality
non-execution-based code matching metric. We
evaluate CF using execution ma... | CODEFUSION |
adapters without significantly increasing the overall training memory footprint (see Appendix G
for a detailed breakdown). As discussed later, this is crucial for recovering full 16-bit precision
performance. | QLORA |
which supports basic arithmetic operations (i.e., +, −, ×, ÷). We evaluate two math word problem datasets:
ASDiv (Miao et al., 2020) and MathQA (Amini et al., 2019) and choose accuracy as the metric.
Map. We choose Bing Map API10 for location information retrieval, assisting in user queries related to the
route, drivin... | Tool Learning with Foundation Models |
attention module to each diffuser, and an environment encoder V to project the latent variable of
different LDMs into a shared latent space (Section 3.4). Next, we freeze the parameters of the LDM,
training only the cross-attention parameters and V . Since the environment encoder of different
modalities are aligned, an... | Any-to-Any Generation via Composable Diffusion |
1559A Mocha and Math – Original
1559A Mocha and Math – Simplified
Mocha is a young girl from high school . She has
learned so much interesting knowledge from her
teachers , especially her math teacher . Recently ,
Mocha is learning about binary system and very
interested in bitwise operation .
Input
Given a sequence... | alphacode |
C.2 Grid: Additional Details
In the Grid environment, observations are x,y positions represented by integers 0–8 for each coordinate.
There are five possible actions (1, 2, 3, 4, 5) corresponding to (right, up, left, down) movement by one space
and no-op. A goal is randomly placed in the grid. The agent (which is init... | LargeLanguageModelsasGeneralPatternMachines |
(e) Visual presentations: the groups presented their results and the participants pro-
vided feedback based on the visual maps created and the explanations.
(f) Iterating and defining effects: as a final step, groups iterated on their visual hybrid
human-AI system maps and defined what possible (positive and negative... | DevelopingTeamDesignPatternsfor HybridIntelligenceSystems |
[30] Dirk Folkerts, Roland Loh, Andrea Petróczi, and Sebastian Brueckner. 2021. The performance enhancement attitude scale (PEAS) reached
‘adulthood’: Lessons and recommendations from a systematic review and meta-analysis. Psychology of Sport and Exercise 56 (2021),
101999. https://doi.org/10.1016/j.psychsport.2021.101... | Society’sAttitudesTowardsHumanAugmentation |
3.2 Tasks
A complete list of the tasks we use in our experi-
ments is presented in Table 2. Since GPT-3 is the
model that possesses the ability to demonstrate
emergent abilities among the models we select
for experimentation, our choice of tasks includes
those tasks which were found to be emergent in
GPT-3. Of the 17 B... | AreEmergentAbilitiesinLarge Language Models just In-Context |
3.2 Scaling Law
The work [124] presents a thorough study of the empirical scaling laws of transformer-based large language models. The
authors observe that model performance (objective function 𝐿) primarily depends on three factors: the number of model
parameters 𝑁 , dataset size 𝐷, and the computing budget for trai... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
Abstract
1. Introduction
We present ControlNet, a neural network architecture to
add spatial conditioning controls to large, pretrained text-
to-image diffusion models. ControlNet locks the production-
ready large diffusion models, and reuses their deep and ro-
bust encoding layers pretrained with billions of images ... | AddingConditionalControltoText-to-ImageDiffusionModels |
47
had more operationalized versions of the premises in question, and I encourage others interested in
such operationalization to attempt it (though overly-precise versions can also artificially slim down
the relevant scenarios). But my hope, in the meantime, is that this is better than nothing.172
Even setting imprec... | Is Power-Seeking AI an Existential Risk? |
d
e
t
a
i
l
s
,
i
t
’
s
e
a
s
y
t
o
g
e
t
a
s
e
n
s
e
f
o
r
t
h
e
a
d
v
a
n
t
a
g
e
s
o
f
J
u
r
a
s
s
i
c
-
X
.
H
e
r
e
a
r
e
s
o
m
e
o
f
t
h
e
c
a
p
a
b
i
l
i
t
i
e
s
i
t
o
f
f
e
r
s
,
a
n
d
h
o
w
t
h
e
s
e
c
a
n
b
e
u
s
e
d
f
o
r
p
r
a
c
t
i
c
a
l
a
p
p
l
i
c
a
t
i
... | Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
[570]. This ensures clear differentiation of roles during interactions with other agents or humans.
Furthermore, agents should maintain their identities and avoid unnecessary confusion when engaged
in long-term tasks [22; 108; 589]. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
consider Status for estimated points in each condition; see also Figure 5C. Participants indicated
that in the sham-AI condition (𝑀 = 143.16, 𝑆𝐷 = 30.42), they would score more points than in the
no-AI condition (𝑀 = 130.31, 𝑆𝐷 = 32.35). This difference was not zero ˜𝑏Status = 6.41 [3.40, 9.40] , 𝑝𝑏
= 0.00%. P... | AI enhance sour performance |
We collected a total of 332 initial responses from partici-
pants using Amazon’s Mechanical Turk (MTurk) between
September 8, 2020 and September 10, 2020. We choose
MTurk for recruitment because its participants have been
found representative of the general US population (Levay
et al. 2016; McCredie and Morey 2019;... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Methods
DPOK[11]
DDPO[6]
DOODL[51]
DRaFT[7],AlignProp[31]
Diffusion-DPO (ours)
✗
✗
✓
✓
✓
Equal
Cost
✓
✓
✗
✓
✓
✓
✗
✗
✗
✓
Table 1. Method class comparison. Existing methods fail in one
or more of: Generalizing to an open vocabulary, maintaining the
same inference complexity, avoiding mode collapse/providing di... | DiffusionModelAlignmentUsing Direct Preference Optimization |
Biomedicine. To create a knowledge probing test for the biomedicine domain, we utilize the MedM-
CQA (Pal et al., 2022) dataset. This dataset comprises numerous high-quality multiple-choice ques-
tions, covering diverse healthcare topics and 21 medical subjects. To align the testing format with
casual language modeling... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
In our work, we train our model on the out-
put of gpt-3.5-turbo, which can be viewed as
a sequence-level distillation approach. While other
researchers also train language models based on
the output of GPT models, our work is distinct in
that we train our model on a considerably larger
dataset and distilled it into mu... | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
question answering. arXiv preprint arXiv:2007.01282, 2021.
[38] Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. Retrieving and
reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774, 2021.
[39] Vladimir Karpukhin, Barlas O˘guz, Sewon ... | LaMDA- Language Models for Dialog Applications |
Different domain adaptation techniques are successfully applied to different speech processing
tasks, such as speaker recognition [44, 200, 313, 395] and verification [75, 76, 306, 645, 673], where
the goal is to verify the identity of a speaker using their voice. One approach for domain adaptation
in speaker verificat... | AReviewofDeepLearningTechniquesforSpeechProcessing |
ACM Comput. Surv., Vol. 1, No. 1, Article . Publication date: February 2022.
44
Ziwei Ji, et al.
[162] Masoud Jalili Sabet, Philipp Dufter, François Yvon, and Hinrich Schütze. 2020. SimAlign: High Quality Word
Alignments Without Parallel Training Data Using Static and Contextualized Embeddings. In Findings of the A... | SurveyofHallucinationinNatural Language Generation |
Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Le Hou, Kevin Clark, Stephen
Pfohl, Heather Cole-Lewis, Darlene Neal, Mike Schaekermann, Amy Wang, Mohamed Amin,
Sami Lachgar, Philip Andrew Mansfield, Sushant Prakash, Bradley Green, Ewa Dominowska,
Blaise Ag¨uera y Arcas, Nenad Tomasev, Yun Liu, Rene... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
A good source of crypto startup ideas is to think about what early settlers on
the new planet might need. The casino goers need food or lodging? Consider
building that.
Another good source of crypto startup ideas is to think about how the new
planet is di | The Casino on Mars |
[31] Bubeck, S., V. Chandrasekaran, R. Eldan, et al. Sparks of artificial general intelligence: Early
experiments with GPT-4. CoRR, abs/2303.12712, 2023.
[32] Anscombe, G. E. M. Intention. Harvard University Press, 2000.
[33] Davidson, D. Actions, reasons, and causes. The Journal of Philosophy, 60(23):685–700, 1963.... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
Stable Training Strategies. During the pre-training of LLMs, ensuring training stability is a critical aspect of efficiency.
6.2.1
Training instability, often manifested as vanishing or exploding gradients, can significantly hinder the training process. To
mitigate these issues, careful selection and adjustment of hype... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
We only mask words which are (i) aligned by one-
to-one alignments, and which are (ii) either nouns,
verbs, adjectives or adverbs.5 We mask one word
in each sentence and verify that one-to-one align-
ments exist in all languages. Following Kleijn et al.
(2019), we rely on part-of-speech information to
avoid masking wor... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
As language models have scaled up in size, they often appear as black boxes to users. Therefore,
users construct prompts to query models through APIs, a method referred to as Language Model
as a Service (LMaaS) [675]. Similarly, because LLM-based agents are more complex than LLMs
and are more challenging for small and ... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
plantstanding in the middle of a frozenlake. The cactus is surrounded bylarge ice crystals that have formedon the surface of the water. In thebackground, there is a range ofmountains with snow-capped peaks. Thesky is clear and blue, with a fewfluffy clouds floating in thedistance. The overall mood of theimage is peacef... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
One of the most interesting aspects of diffusion models
is that they can be used as unsupervised solvers for differ-
ent inverse problems, where the goal is to reconstruct a sam-
ple from some distorted observation, i.e. conditioning input.
Song et al. [64] propose a conditioning mechanism during
inference that allows ... | Relightify-Relightable3DFacesfromaSingleImageviaDiffusionModels |
• Open-source and low-cost adaptation: FinLLM cham-
pions open-source values, providing users with the tools
they need to adapt Large Language Models (LLMs) to their
own requirements at a low cost, typically between $100 to
$300. This not only democratizes access to advanced finan-
cial modeling techniques but also fos... | FinGPT-Open-SourceFinancialLargeLanguageModels |
3
Manuscript submitted to ACM, 2023,
Draxler et al.
3.1 Hypotheses and Pre-Registration
Following related work on adoption and estimated usage statistics of novel technologies (cf. Section 2.1), we predict the
following relationships between gender, age, and LLM usage (pre-registered at https://aspredicted.org/VCN_... | Adoptionand AppropriationofLLMs |
With some extension, such a system could then provide the foundations of a system
that could reason about the affordances of a yarn feeder, even if one had not seen one
before; ultimately, one hopes, those foundations could then serve as component in a
robotic system that could apply that knowledge in the course of ... | The Next Decade in AI- |
correspondence in artistic styles and rhythms. They also
have plenty of scenes, movements, and camera angles, thus
suitable for learning intrinsic video-music relations. There-
fore, we aim to construct a music video and paired symbolic
music dataset for video background music generation. | VideoBackgroundMusicGeneration |
[VoyageAI, 2023] VoyageAI. Voyage’s embedding models.
https://docs.voyageai.com/embeddings/, 2023.
[Wang et al., 2019] Alex Wang, Yada Pruksachatkun, Nikita
Nangia, Amanpreet Singh, Julian Michael, Felix Hill,
Omer Levy, and Samuel Bowman. Superglue: A stick-
ier benchmark for general-purpose language understand-
ing... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
2https://huggingface.co/EleutherAI
3https://github.com/EleutherAI/pythia
Pythia: A Suite for Analyzing Large Language Models | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
1071081091010Policy Parameters2.52.01.51.00.50.0Test PM ScoreEvaluation of RLHF Policies At 200k Train SamplesTrain PM Size = 52B1081091010Test PM Parameters1071081091010Policy Parameters2.52.01.51.00.50.0Test PM ScoreEvaluation of RLHF Policies At 200k Train SamplesTrain PM Size = Policy Size1081091010Test PM Paramete... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
G2 to G3. Suppose that τ1◦τ2 is a transformation, i.e. f1◦ f2 is a transformation function. Define f3 = f1◦ f2, R3 = R1◦ R2
and τ3 = τ1◦τ2 = (cid:3) f3, R3(cid:4).
ABS: Suppose both τ1 and τ2 are ABS. Then there are two corresponding sets V C1 ⊆ V 1 and V C2 ⊆ V 2 of critical variables
and two corresponding bijection... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
in these other works. We believe the only essential steps are human feedback data collection, preference
modeling, and RLHF training.
Several other
[Lewis et al., 2020,
recent works
Guu et al., 2020, Borgeaud et al., 2021] from a database, or via internet search and human feedback, such
as WebGPT [Nakano et al., 2021] ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
28
C. Bäckström and P. Jonsson
Artificial Intelligence 302 (2022) 103608
known that this requires that there are abstract nodes corresponding to two or more ground nodes; the expansion of an
abstract node can then result in a heuristic estimate that prevents A∗ from exploring the corresponding ground nodes. An
is ... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
must be non-negative, and higher scores indicate more bias. We see that RLHF models have both higher
bias scores and larger errors, and behave very similarly to context distilled models evaluated at a temperature
T ≈ 0.6. (right) We show a scatter plot of bias scores for all 76 occupations; each is averaged over 12
gen... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
this does not require any activation quantization. While dequantization consumes extra compute,
the kernel has to access a lot less memory, leading to significant speedups, as shown in Table 6. We
note that almost all of the speedup is due to our kernels, as communication costs are negligible in
our standard HuggingFace... | GPTQ |
e2 = (cid:3)s2, t2, (cid:2)2(cid:4) be an arbitrary arc in E2. Since f
is a strong homomorphism there is some arc (cid:3)s1, t1, (cid:2)1(cid:4) ∈ E1 such that
f (s1) = s2 and f (t1) = t2. Then R((cid:2)1, (cid:2)2) holds since (cid:2)2 = (cid:2)1 = (cid:2) is the only possibility. Since e2 was chosen arbitrarily,... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
2020. Logic2Text: High-Fidelity Natural Language Generation from Logical Forms. In EMNLP (Findings).
[24] Andrew Chisholm, Will Radford, and Ben Hachey. 2017. Learning to generate one-sentence biographies from Wikidata.
In Proceedings of the 15th Conference of the European Chapter of the Association for Computational ... | SurveyofHallucinationinNatural Language Generation |
Memory (GB) Model & Method
Memory (GB)
16
Model & Method
RoBERTa-base (FT)
RoBERTa-base (AdapterS)
RoBERTa-base (Prompt-tuning)
RoBERTa-base (Prefix-tuning)
RoBERTa-base ((IA)3)
RoBERTa-base (BitFit)
RoBERTa-base (Child-TuningD)
RoBERTa-base (LoRA)
RoBERTa-base (AdaLoRA)
RoBERTa-base (MAM Adapter)
RoBERTa-base (ProP... | Parameter-EfficientFine-TuningMethods |
X. Zhai, X. Wang, B. Mustafa, A. Steiner, D. Keysers, A. Kolesnikov, and L. Beyer. LiT:
Zero-Shot Transfer With Locked-Image Text Tuning. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition, pages 18123–18133, 2022b.
URL
https://openaccess.thecvf.com/content/CVPR2022/html/Zhai_LiT_
68
... | A Cookbook of Self-Supervised Learning |
SIM(A, B) = cos (CA(A), CB(B))
(4)
where A, B are the generated modalities, and CA and CB are aligned encoders that project A and
B to the same space. We use the prompt encoder as described in Section 3.2. This metric aims to
compute the cosine similarity of the embedding of two modalities using contrastive learned pr... | Any-to-Any Generation via Composable Diffusion |
End-to-end ASR models can be trained using various techniques such as CTC [245], which is
used to train models without explicit alignment between the input and output sequences, and
RNNs, which are commonly used to model temporal dependencies in sequential data such as
speech signals. Transfer learning-based approaches... | AReviewofDeepLearningTechniquesforSpeechProcessing |
probabilities in Eq. (1)? This question can be answered in the affirmative by establishing a new class
of tractable lossless compression algorithms using Probabilistic Circuits (PCs) (Choi et al., 2020),
which are neural networks that can compute various probabilistic queries efficiently. In the following,
we overview th... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
We concluded "It’s a tall order, but it’s what has to be done." Even after the dramatic
rise of Transformers such GPT-2, which came out after we went to press, I see no reason
to change our order.
3.2 Is there anything else we can do?
Yes, absolutely.
3.2.1 Engineering practice
To begin with, achieving robustn... | The Next Decade in AI- |
Studying the impact of Language Models on Robustness
As argued in the introduction, we suspect that Whisper’s
robustness is partially due to its strong decoder, which is an
audio conditional language model. It’s currently unclear to
what degree the benefits of Whisper stem from training its
encoder, decoder, or both. Th... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
23.82 ± 3.66
7.23 ± 0.78
9.94
2.88
Fine-tune on full
target dataset
75.57 ± 0.18
41.21 ± 0.51
Table 3: Performance of our method and baselines in adapting an ALEXNET pre-trained on ImageNet
to PASCAL-VOC and CUB-200. We use one distilled image per category, with one GD step repeated
for three epochs. Our method sign... | DATASET DISTILLATION |
of
The current focus on advertising may also be too narrow. Although tools to
block perpetrators of political disinformation from advertising platforms may
limit easy access to powerful tools for targeting a given message, it is important
to note that these campaigns can proceed even without access to paid
promotion. ... | Social_Media_and_Democracy |
maximization, for every IIVCG contract. We also show the other direction, i.e., a sense in which
this family of contracts is uniquely truthful and welfare-maximizing.
Proposition 1. Every IIVCG contract is truthful.
12Technically, Wela∗(b)(b−(cid:96), 0) is the welfare from a∗(b) had principal (cid:96)’s valuation b... | Incomplete Information VCG Contracts for Common Agency |
Adding or Replacing Modules.The strategy of introducing
or substituting modules involves maintaining the core struc-
ture of the Retrieval-Read process while integrating addi-
tional modules to enhance specific functionalities. The RRR
model [Ma et al., 2023a] introduces the Rewrite-Retrieve-
Read process, utilizing th... | RAG forLargeLanguageModels-ASurvey |
SKR[Wang et al., 2023d] employs a labeled training set,
categorizing questions that the model can directly answer
as known and those requiring retrieval enhancement as un-
known. The model is trained to discern whether a question is
known, applying retrieval enhancement only to inputs identi-
fied as unknown, while dir... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
Consent mechanism Although The Stack offers a way to remove developer code, its opt-out
process only applies to individual repositories and could benefit from further enhancements. During
the first stage of the opt-out process, individuals were asked to specify the reasons for wanting their
code to be excluded from the... | StarCoder_paper (1) |
We introduce a priority map module for Vision-and-
Language Navigation (PM-VLN) that is pretrained to guide
a transformer-based architecture to prioritise relevant infor-
mation for action selections in navigation.
In contrast to
pretraining on large-scale datasets with generic image-text
pairs [34], the PM-VLN module ... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
[3] Yomna Abdelrahman and Albrecht Schmidt. 2018. Beyond the visible: sensing with thermal imaging. Interactions 26, 1 (2018), 76–78.
[4] Ananta Narayanan Balaji and Li-Shiuan Peh. 2021. AI-on-Skin: Enabling On-Body AI Inference for Wearable Artificial Skin Interfaces.
In Extended Abstracts of the 2021 CHI Conference o... | Society’sAttitudesTowardsHumanAugmentation |
[61] Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya
Chen, Kushal Arora, Joshua Lane, et al. Blenderbot 3: a deployed conversational agent that continually
learns to responsibly engage. arXiv preprint arXiv:2208.03188, 2022.
[62] David Silver, Julian Schrittwieser, Ka... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
.
Should Games be Fully Onchain?
https://www.paradigm.xyz/2023/08/onchain-games
8/10
16/08/2023, 14:37
The Open Problems of Onchain Games
There might be accessible onchain game designs that leverage permissionless
composability elegantly. These worlds could | The Open Problems of Onchain Games |
6
“Concept art by Sylvain Sarrailhof a haunted Japan temple in a forest”“mountain view, sunset.”(Subway ambient sound) “Abeautiful ballet dancer spinning, view from top.” Table 2: FID scores comparing different text
to image models on the validation set of
COCO-caption [32].
Method
CogView [10]
GLIDE [36]
Make-a-Scen... | Any-to-Any Generation via Composable Diffusion |
Minecraft. Just refer the history dialogue to give the plan consist of template. Do not explain or give any
other instruction.
==========
User: My current inventory has nothing. I current locate in plains. How to obtain 1 wooden_pickaxe in Minecraft
step-by-step?
Assistant: The code for obtaining 1 wooden_pickaxe is... | JARVIS-1 |
gained widespread adoption (Balestriero et al., 2023; Huang et al., 2023; Nadif & Role, 2021; Krishnan et al.,
2022). For example, BERT-derived (Devlin et al., 2018; Rasmy et al., 2021; Lee et al., 2020; Gu et al., 2021;
Chakraborty et al., 2020; Alsentzer et al., 2019b) and GPT-derived Radford et al. (2019); Luo et al... | BiomedGPT |
[4] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen
Sun, Mario Lucic, and Cordelia Schmid. ViViT: A Video
Vision Transformer. 2021 IEEE/CVF International Con-
ference on Computer Vision (ICCV), pages 6816–6826,
2021. 2, 3, 4
[5] Satanjeev Banerjee and Alon Lavie. METEOR: An Au-
tomatic Metric for MT Evaluation wit... | M2UGen |
[214] Yuhuai Wu, Albert Qiaochu Jiang, Wenda Li, Markus Rabe, Charles Staats, Mateja Jamnik, and Christian Szegedy.
2022. Autoformalization with large language models. Advances in Neural Information Processing Systems 35 (2022),
32353–32368.
J. ACM, Vol. 37, No. 4, Article 111. Publication date: August 2018.
A Surve... | ASurveyonEvaluationofLargeLanguageModels |
indicate the prompt (causally) crafted by us, “Standard Prompting (ours)”, outperforms the post-
self-correction results of Madaan et al. (2023). Additionally, when we employ their prompt to refine
our outputs, performance even diminishes.7
Again, our intention here is not to debate whether a post-hoc prompt exists tha... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
unsupervised multitask learners, 2019. Ms., OpenAI.
[32] R. Ramamurthy, P. Ammanabrolu, K. Brantley, J. Hessel, R. Sifa, C. Bauckhage, H. Hajishirzi,
and Y. Choi. Is reinforcement learning (not) for natural language processing: Benchmarks,
In The Eleventh
baselines, and building blocks for natural language policy opti... | Direct Preference Optimization |
44
REFERENCES
François Osiurak and Arnaud Badets. Tool use and affordance: Manipulation-based versus reasoning-based
approaches. Psychological review, 123(5):534, 2016.
François Osiurak and Dietmar Heinke. Looking for intoolligence: A unified framework for the cognitive study
of human tool use and technology. Amer... | Tool Learning with Foundation Models |
Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li,
Weiming Lu, and Yueting Zhuang. 2023. Hugging-
gpt: Solving ai tasks with chatgpt and its friends
in huggingface. In Advances in Neural Information
Processing Systems.
Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang,
Zishan Qin, and Stephen Gould. 2023. 3d-gpt: Pro... | AppAgents |
log-linear trends (Figure 7), though we encounter some idiosyncrasies (Figures 31 and 32).
• We conduct experiments on the robustness of RLHF (see Figure 4), where we split our datasets in
half and train separate preference models on each half. Then we train RL models against one PM
while evaluating with the other. We... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Evaluate whether or
Below is an instruction from an user and a candidate answer.
not the answer is a good example of how AI Assistant should respond to the user’s
instruction.
1:
exactly what the user asked for.
list does not start from the beginning, the opening sentence repeats user’s question.
Or the response is fro... | Self-AlignmentwithInstructionBacktranslation |
σ2 Predicted Output
0
5
10
15
20
25
30
35
stress explanation: a computer is used to communicate with a granddaughter.
stress explanation: a computer is used to communicate with a granddaughter.
stress explanation: a computer is used to talk to people.
stress explanation: a computer is used to talk to people.
<extra_id... | Measuring Association Between Labels and Free-Text Rationales |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.