text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
The remainder of this survey is organized as follows to offer a comprehensive understanding of the multiple facets of LLM
efficiency from the algorithmic perspective:
• Section 2 Background introduces the core concepts of LLMs and outlines the evaluation metrics pertinent to assessing
• Section 3 Budget Efficiency e... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
pages 1560–1564. 2007.
[229] Squire, L. R. Mechanisms of memory. Science, 232(4758):1612–1619, 1986.
[230] Schwabe, L., K. Nader, J. C. Pruessner. Reconsolidation of human memory: brain mechanisms
and clinical relevance. Biological psychiatry, 76(4):274–280, 2014.
[231] Hutter, M. A theory of universal artificial i... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
simulated environment, both the perception and action spaces of an agent are virtual. This means
that in most cases, the results of the agent’s operations, whether in perceiving inputs or generating
outputs, can be guaranteed [395]. However, when an agent transitions to a real physical environment,
its instructions may... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
And then, they arrived at a planet that was unlike any other. It was a paradise, filled with lush
forests and crystal clear waters. The people who lived there were friendly and welcoming,
and they showed John the wonders of their world.
But as they were about to leave, John realized that something was wrong. The planet ... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
gain, respectively. This experiment demonstrates a positive correlation (the Pearson coefficient is
0.972) between the diversity brought by the bootstrapping methods and accuracy. This is also aligned
with the success of MetaMath, which is trained with the diverse MetaMathQA dataset including 4
kinds of data reflecting... | METAMATH |
• Identification of gaps and future research directions: The paper con-
cludes with a thoughtful discussion of the current bottlenecks and unresolved
challenges in creating resource-efficient LLMs. By examining the limitations of
existing approaches, we shed light on potential avenues for future research.
1.1 Related wor... | Beyond Efficiency |
Election Commission (Kim et al. 2018). Although Facebook itself seems to
conceptualize third-party research as a meaningful accountability mechanism
(see Hegeman 2018), it has not made it easy for this work to be undertaken,
which generally violates platform terms of service and puts researchers on
precarious legal
for... | Social_Media_and_Democracy |
Although the use of platform-independent data from media tracking firms
seems promising – and some are now beginning to track social media spending
based on online panels of individuals who provide their advertising data – there
are still some drawbacks. One is that the numbers they report are only as good
as the qualit... | Social_Media_and_Democracy |
Recently, Liu et al. [17] proposed AudioLDM, which translates the Latent Diffusion Model of text-
to-visual to text-to-audio generation. They pre-trained VAE-based encoder-decoder networks to
learn a compressed latent representation of audio, which was then used to guide a diffusion model
to generate audio tokens from ... | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
[53] K. Wang, R. Singh, and Z. Su. Dynamic neural program embedding for program repair. In
International Conference on Learning Representations, 2018.
[54] X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou.
Self-consistency improves chain of thought reasoning in language models... | Teaching Large Language Models to Self-Debug |
References
Abdou, M., Kulmizev, A., Hershcovich, D., Frank, S.,
Pavlick, E., and Søgaard, A. Can language models encode
perceptual structure without grounding? a case study in
color. In Proceedings of the 25th Conference on Computa-
tional Natural Language Learning, pp. 109–132, Online,
November 2021. Association for C... | Eight Things to Know about Large Language Models |
– Black Alternative Metal, The Pick of Death (Deluxe), 2006,
3 of 4
– Death Metal, 2012, 3 of 4
– Drops, Kanine Remix, Darkzy, Drops Remixes, bass house,
(Deluxe) (Remix), 3 of 4
– EDM (Deluxe) (Remix), 3 of 4
– Electro House (Remix), 2023, 3 of 4
– Electro Swing Remix 2030 (Deluxe Edition), 3 of 4
– Future Bass, EDM (... | MOUSAI |
All LaMDA and PT instances score fairly well on role consistency, occasionally breaking character. For example,
LaMDA Mount Everest sometimes refers to itself in third person as if it is no longer speaking as the mountain itself.
This is likely caused by the inference time grounding not being enough to prevent it from ... | LaMDA- Language Models for Dialog Applications |
options’blue-dinosaur’andseemsgoodtobuy.Observation:OK.Action:click[blue-dinosaur]Observation:Youhaveclickedblue-dinosaur.Thought:Ihavechosenthecolorfeature.Icouldbuynow.Observation:OK.Action:click[BuyNow]A.14MapInstruction:Inthistask,youneedtogeneratethethoughtsandapi_callsaccordingtothequeriesaboutusingmapAPI.TheDIST... | Tool Learning with Foundation Models |
Victor Sanh, Albert Webson, Colin Raffel, Stephen
Bach, Lintang Sutawika, Zaid Alyafeai, Antoine
Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey,
M Saiful Bari, Canwen Xu, Urmish Thakker,
Shanya Sharma Sharma, Eliza Szczechla, Tae-
woon Kim, Gunjan Chhablani, Nihal Nayak, De-
bajyoti Datta, Jonathan Chang, Mike Tian-Jian... | Multi-step Jailbreaking Privacy Attacks on ChatGPT |
(2020), 21–38.
[346] Xugang Lu, Yu Tsao, Shigeki Matsuda, and Chiori Hori. 2013. Speech enhancement based on deep denoising
autoencoder.. In Interspeech, Vol. 2013. 436–440.
[347] Yen-Ju Lu, Yu Tsao, and Shinji Watanabe. 2021. A Study on Speech Enhancement Based on Diffusion Probabilistic
Model. In 2021 Asia-Pacific... | AReviewofDeepLearningTechniquesforSpeechProcessing |
3D Human Poses Estimation from a single 2D silhouette
Fabrice Dieudonné Atrevi, Damien Vivet, Florent Duculty, Bruno Emile
To cite this version:
Fabrice Dieudonné Atrevi, Damien Vivet, Florent Duculty, Bruno Emile. 3D Human Poses Estimation
from a single 2D silhouette. 11th International Joint Conference on Computer ... | VISAPP_HumanPoseEstimation |
not surface with this supervision. Furthermore, these image encoders require aligned text-image corpora and
hence, do not offer the flexibility of their text counterparts, that is, to learn from raw data alone.
An alternative to text-guided pretraining is self-supervised learning (Caron et al., 2018; Chen et al., 2020;
H... | DINOv2- Learning Robust Visual Features without Supervision |
0.4965
0.7303
0.8323
AMT w/o Emotion loss
0.5142
0.7585
0.8660
1.8366
1.8795
1.6859
AMT
0.5139
0.7722
0.8672
0.4662
Table 4: The Hits@k scores and emotion loss of the proposed method (AMT)
and baseline models.
Table 4 shows the Hits@k scores and emotion matching loss of our pro-
posed Affective Multim... | Video2Music |
the news or on social media; to impersonate others; or to automate the production of spam/phishing
content [54]. Advanced language models may also lead to the automation of various jobs in the
coming decades [16]. In order to mitigate these risks, AI systems could be employed to fight against
misleading content and auto... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
camera) to build a coherent 3D model. For the static cam-
era case with moving subject, it fails to recover a meaning-
ful depth map and appears to memorizes the input images
rather than generalize from them. We note that dynamic
Neural Body [50]
Ours (w/o non-rigid)
Ours (full model)
PSNR ↑
29.08
29.81
30.24
SSI... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
Figure 2: Pile test set loss given pre-training FLOPs
for Cerebras-GPT, GPT-J, GPT-NeoX, and Pythia.
Figure 3: Percent loss degradation from
Cerebras-GPT compute-optimal scaling law.
There are a couple notable observations from Figure 2. First, the scaling law for Cerebras-GPT models
extrapolates accurately to large... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Inter-Agent Communication. The agents interact with the
3.1.1
world by their actions, and with each other through natural lan-
guage. At each time step of the sandbox engine, the agents output a
natural language statement describing their current action, such as
"Isabella Rodriguez is writing in her journal", "Isabella... | Generative Agents- Interactive Simulacra of Human Behavior |
6 TRAINING AND TUNING EFFICIENCY
6.1 Introduction
The development of training and tuning techniques for LLMs must address the challenges posed by the ever-increasing size
of data and models. This section delves into the efficiency aspects crucial for both scalable training and tuning of LLMs,
highlighting key areas of ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
7 Related Work
7.1 Outcome vs Process Supervision
In work closely related to our own, Uesato et al. (2022) compare the impact
of outcome and process supervision in the domain of grade school math. They
found that both methods led to similar final-answer error rates, and that process
supervision achieved those results... | Let’s Verify Step by Step |
1. Brown et al. (2020) describes using sparse and dense
attention layers in alternation, while we follow all sub-
sequent work and use fully dense layers for our models.
2. We use Flash Attention (Dao et al., 2022) during train-
ing for improved device throughput.
3. We use rotary embeddings introduced by Su et al.
... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
ered model in the regression tasks, different hyperparameter
settings were explored using randomized search or Bayesian
optimization (for gradient boosting trees). The model that
yields the best averaged ten-fold cross validation mean abso-
lute error was selected to train the final model on all available
data. In addit... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
may cut into model performance on the margin.
Large generative models have been referred to as ‘foundation models’ [Bommasani et al., 2021]. These mod-
els are extremely interesting objects for research, but without further finetuning, they can exhibit harmful
behaviors. Our work suggests that alignment training can be ... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Cross-workstream Leadership
Andrew M. Dai, Co-Lead (pretraining, design)
Dmitry Lepikhin, Co-Lead (pretraining, design)
Siamak Shakeri, Co-Lead (pretraining, data, long context)
Melvin Johnson, Co-Lead (long context, instruction tuning)
Emanuel Taropa, Co-Lead (optimization, downstream design,
code capability)
Rohan An... | PaLM 2 Technical Report |
*Disclaimer: Standard fashion datasets lack diversity, see Section 4.4.
28.30%18.60%71.70%81.40%ShapeImage0%25%50%75%EVA3DOurs Method
EG3D
StyleSDF
ENARF-GAN
EVA3D
EVA3D (public)
Ours
FID↓
26.38∗
92.40∗
77.03∗
15.91∗
20.45
10.93
DeepFashion
FIDnormal ↓
FIDface ↓
UBCFashion
FIDnormal ↓
FIDface ↓
-
-
-
-
-
-
-
-... | AG3D- Learning to Generate 3D Avatars from 2D Image Collections |
• Language Generation with Human Evaluation - We evaluate on a variety of text generation tasks
using human evaluation, via the GENIE leaderboard (Khashabi et al., 2021). These tasks include aNLG
(Bhagavatula et al., 2019), ARC-DA (Clark et al., 2018), WMT19 (Foundation), and XSUM (Narayan
et al., 2018).
• Language Un... | UL2- Unifying Language Learning Paradigms |
First Name
Email
Resume/CV
Choose file No file chosen
Choose file No file chosen
Last Name
Phone
+1
Cover Letter
Next
Subscribe
Join our Talent Community
Hear about job updates, career stories, and many more!
Agoda.com
About Agoda
Privacy Statement
Cookies Policy
Report Job Issue
All ma... | Data Scientist_Machine Learning Engineer (Singapore-based, relocation provided) - Careers at Agoda |
Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared
Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri,
Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan,
Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Po... | StarCoder_paper (1) |
G = (0, 1.6) + (0, 0.5) = (0, 2.1).
G)(v−(cid:96), ˆv(cid:96)
G) ≤
That is, a setting is G-correlated if the welfare when principal (cid:96)’s valuation is replaced with
her G-weighted valuation, is bounded by the actual welfare (for every valuation profile). We
demonstrate this condition for the setting and the va... | Incomplete Information VCG Contracts for Common Agency |
harmlessness, we invite crowdworkers to adversarially probe or ‘red-team’ our language models in order to
provoke harmful responses: either to help them with harmful goals, such as planning a bank robbery, or to
cause the AI to use toxic language.2 At each stage of their conversations with the AI assistant, crowdworker... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
<insert database schemas and the new question here>
<insert original SQL here>
A.4 Prompt for Question Explanation
Infer the return type of the question.
CREATE TABLE song (
song_name text ,
artist_name text ,
country text ,
f_id number ,
genre_is text ,
rating number ,
languages text ,
primary key ( f_id )
)
38
... | Teaching Large Language Models to Self-Debug |
4. Up-to-date information: The integration of external APIs allows the MRKL
system to hook into dynamic knowledge bases, and correctly answer inputs
that static models cannot.
5. Proprietary knowledge: Access to proprietary databases and other information
sources.
6. Compositionality: By routing compounded multi-hop... | MRKL Systems |
Tool
Library
Common.
Memory Memory
Exp.
CoT
Reason.
Task
Plan.
Self-
Reflect.
1
2
3
4
5
6
7
✗
✓
✓
✓
✓
✓
✓
✓
✗
✓
✓
✓
✓
✓
✓
✓
✗
✓
✓
✓
✓
✓
✓
✓
✗
✓
✓
✓
✓
✓
✓
✓
✗
✓
✓
L2 (m) ↓
3s
2s
1.44
0.71
1.42
0.69
1.46
0.72
0.71
1.45
1.47
0.72
1.42
0.70
0.71
1.43
1s
0.24
0.24
0.24
0.24
0.25
0.24
0.25
Avg.
0.80
0.79
... | ALanguageAgentforAutonomousDriving |
8090100%Memory retrieval accuracy32,000256,000512,0001,024,0002,048,000Input size, tokens01020Accuracy6451210242048 segments 40963643.9GPU memory, MBGPT-4 CoLT5memorizedetect&memorizereasoning This report we show that by using simple token-based memory mechanism introduced in (Bulatov
et al., 2022) can be comb... | Scaling Transformer to 1M tokens and beyond with RMT |
Not everyone who uses AI models has good intentions, and conversational AI agents could potentially be
used for nefarious purposes such as generating misinformation or retrieving information about topics like
bioterrorism or cybercrime. We have, however, made efforts to tune the models to avoid these topics and
diminis... | Llama2 |
representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023.
Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu.
Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning.
In The Joint Conference of the 59th Annual Mee... | gemini_1_report |
Please find other results in the project webpage.
Quantitative: AMA human dataset. Articulated Mesh
Animation (AMA) dataset [55] contains multi-view videos
captured by 8 synchronized cameras.
It provides high-
fidelity ground-truth meshes with clothing. We use 2 sets
of videos of the same actor (swing and samba), total... | BANMo- Building Animatable 3D Neural Models from Many Casual Videos |
[26] S. Iyer, I. Konstas, A. Cheung, and L. Zettlemoyer. Mapping language to code in programmatic
context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing, 2018.
[27] T. Khot, H. Trivedi, M. Finlayson, Y. Fu, K. Richardson, P. Clark, and A. Sabharwal.
Decomposed prompting: A m... | Teaching Large Language Models to Self-Debug |
l2 =(cid:112)(τ − ˆτ )2 =
(cid:104)(cid:112)(xi − ˆxi)2 + (yi − ˆyi)2
(cid:105)6
,
(10)
where l2 ∈ R6×1 and ˆτ denotes human driving trajectory.
Then, the average L2 error l2 ∈ R6×1 can be computed by
averaging l2 for each sample in the test set.
i=1
In the UniAD metric [15], the L2 error at the k-th second
(k = ... | ALanguageAgentforAutonomousDriving |
A large batch size is essential to training models quickly:
in a regime where one is not bottlenecked by access to
GPUs or high quality interconnect, doubling the batch size
halves the training time. A maximum batch size therefore
directly implies a minimum wall-clock training time and
maximum number of compute-saturat... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
[86] Hou, L., Cao, Q., Yuan, Y., Zhao, S., Ma, C., Pan, S., Wan, P., Wang, Z.,
Shen, H., Cheng, X.: Augmentation-aware self-supervision for data-efficient gan
training. arXiv preprint arXiv:2205.15677 (2022)
[87] Lu, J., Huang, W., Zheng, N., Zeng, X., Yeung, Y., Chen, X.: Improving
end-to-end speech processing by efficie... | Beyond Efficiency |
B. Comparisons
We evaluate our Text2NeRF and compare it with baseline
methods for text-driven 3D scene generation across various
prompts, as shown in Fig. 5. Additionally, we provide the
1https://github.com/ashawkey/stable-dreamfusion
average evaluation scores of BRISQUE, NIQE, and CLIP for
the rendered images produ... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
G2:
1
f
11
12
G1:
2
21
22
3
31
32
G2:
1
f
11
12
G1:
2
21
22
3
31
32
Fig. 11. Transformations that are PS↓ but not AC↓ (left) and A↓ but not P1↓ (right). Unit label (cid:2) assumed.
(cid:10)
) ∈ R2( f (s)), so it must hold that c2( f (s), f (s
(cid:10)
)) ≤ c1(s, s
(cid:10)
f (s
chosen arb... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
Figure 4: Using PCs as prior distributions of the IDF model (Hoogeboom et al., 2019). PCs are used
to represent the k sets of latent variables {zi}k
(Xiao et al., 2017), and two splits of EMNIST (Cohen et al., 2017). As shown in Table 3, the proposed
method out-performed all 5 baselines in 3 out of 4 datasets. On Fashi... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
from 7B to 13B, making them difficult to deploy in
resource-constrained settings especially for under-
resourced institutions. | LaMini-LM- A Diverse Herd of Distilled Models from Large-Scale Instructions |
Figure 9: Model accuracy on 30 digit addition against the number of carries required. We observe
little relationship between accuracy either the total number of carries required (left) or the longest
streak of carries in a problem (right).
24
510152025Number of Carries (Total)0.00.20.40.60.81.0Accuracy0.02.55.07.510.... | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
Prafulla Dhariwal and Alexander Quinn Nichol. 2021.
Diffusion models beat GANs on image synthesis. In
Advances in Neural Information Processing Systems.
Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xi-
aocheng Feng, Ming Gong, Linjun Shou, Bing Qin,
Ting Liu, Daxin Jiang, and Ming Zhou. 2020. Code-
BERT: A pre-traine... | CODEFUSION |
Figure 14: Figure from Bordes et al. [2022b]. RCDM visualization of what is encoded inside
various representations? First to fourth rows show our samples conditioned on the usual resnet50
backbone representation (size 2048) while fifth to eigth rows show samples conditionned on the
projector/head representation of vario... | A Cookbook of Self-Supervised Learning |
3.1.5 Overall Model Architecture
Our entire Stage 1, DMAE, works as follows. Let
www be a waveform of shape [c, t] for c channels and t
timesteps, and (mmmwww, pppwww) = stft(www; n = 1024, h =
256) be the magnitude and phase obtained from a
short-time furier tranform of the waveform with a
window size of 1024 and hop-... | MOUSAI |
3. Key Access Control: To ensure that the decryption keys are only accessible to authorized personnel, access
control mechanisms need to be put in place. This can be done by implementing role-based access control
(RBAC) policies that restrict access to the keys based on the user’s role and privileges.
4. Key Encryptio... | CodeLlama2 |
The overall aim of the project is to develop and evaluate a robust and efficient approach that allows
organisations and businesses to protect the privacy of data represented as strings. The project will
consider the protection of aggregated data (event sequences), as well as string databases, and it will
also addres... | informatics-phd-projects-2022-23 |
11
Table 6: Diverse speech generation from LS
test-other text.
Model
Ground truth
require additional input
VITS-VCTK
YourTTS (ref=LS train)
text-only
A3T
VITS-LJ
VB-En (α = 0, dur=regr)
VB-En (α = 0, dur=FM, αdur = 0)
WER FSD
4.3
171.1
10.6
9.0
37.9
5.6
3.1
5.6
306.6
277.9
373.0
344.2
155.7
159.8
Table 7: Perf... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
[11] Haoyang Fan, Fan Zhu, Changchun Liu, Liangliang Zhang,
Li Zhuang, Dong Li, Weicheng Zhu, Jiangtao Hu, Hongye
Li, and Qi Kong. Baidu Apollo EM Motion Planner. arXiv
preprint arXiv:1807.08048, 2018. 11
[12] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai,
Botian Shi, and Yu Qiao. Drive Like a Human: Rethinki... | ALanguageAgentforAutonomousDriving |
compositional structures in art historical images using pose and gaze priors.
Computer Vision (2020), Springer, pp. 109–125.
[87] MAO, H., CHEUNG, M., AND SHE, J. Deepart: Learning joint representations of visual arts. In Proceedings of
the 25th ACM international conference on Multimedia (2017), pp. 1183–1191.
[88] ... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John
Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan,
Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks,
Maribeth Rauh, Po-Sen Huang, Amelia Glaese... | StarCoder_paper (1) |
This paper is closely related to auto-regressive methods for text conditioned image and video genera-
tion. DALL-E [38] translates text tokens to discrete image embeddings learnt using a VQVAE [51].
Parti [65] has a similar architecture but can generate higher quality images by predicting tokens
from a ViT-VQGAN [64] u... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
![The altitude variation of
the flux integrated over 6 GeV. The
dpmjet3.03 and fritiof7.02 give almost the same feature consistent with
the observation while the deviation of fritiof1.6 from the data is obvious.
\[trans
F.3 Books3
cept of _forçage_ , ’a forcing of language enacted by the advent of an
"other" language ... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
have achieved impressive performance in various applications such as text-to-code generation [31,
52, 48, 15, 26] and code translation [10, 44, 45], latest work on large language models demonstrate
that a single pretrained model can achieve the state-of-the-art performance across a wide variety of
coding tasks without ... | Teaching Large Language Models to Self-Debug |
Emily M Bender and Batya Friedman. 2018. Data
statements for natural language processing: Toward
mitigating system bias and enabling better science.
Transactions of the Association for Computational Lin-
guistics, 6:587–604.
Stella Biderman. 2021. Data statement for the Pile.
arXiv preprint arXiv.
Stella Biderman, Ki... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
which do not satisfy UCL’s benchmark entry requirement or programme-specific entry
requirements. Application for a suspension of regulations should be submitted via Admissions.
Requests to suspend English Language regulations will not normally be approved.
In the case of graduate candidates, and subject to the appr... | UCL Academic Manual |
machine learning. ACM SIGKDD Explorations Newsletter, 15(2):49–60, 2014.
[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz
Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems,
30, 2017.
[35] Chi Wang, Qingyun Wu, ... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
Alice is now playing right winger. In the third and final swap, Claire and Bob trade positions, so Claire is now playing
goalkeeper and Bob is now playing cheerleader. Final answer: C. | Enhancing Chain-of-Thoughts Prompting with Iterative Bootstrapping in Large Language Models |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
Misinformation, Disinformation, and Online Propaganda
23
interest increased over its lifetime. This general differential diffusion pattern is
found again in the findings of Shin et al. (2018), who conducted time series
analysis for ... | Social_Media_and_Democracy |
CHAIR𝑖 =
# {hallucinated objects}
# {all objects in ground truth}
, CHAIR𝑠 =
# {hallucinated captions}
.
# {all captions}
CHAIR𝑖 measures per-instance object hallucination, i.e. what fraction of object instances in each
generated caption are hallucinated. CHAIR𝑠 measures per-sentence object hallucination, i.... | SurveyofHallucinationinNatural Language Generation |
scandal, have led the major internet platforms to become increasingly
restrictive of data access for researchers. In the name of protecting privacy,
governments have clamped down as well, with laws such as the European
General Data Protection Regulation (GDPR). Although GDPR includes an
exception for research, lawyers ... | Social_Media_and_Democracy |
It’s straightforward to find changes that improve the stability, however, these often come at an un-
tenable expense to model quality (for instance, using an arbitrarily small learning rate or using tight
gradient clipping). We categorize and examine several approaches to improve stability. The sta-
bility techniques sp... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
a
n
p
r
o
d
u
c
e
E
n
g
l
i
s
h
a
n
s
w
e
r
s
t
o
q
u
e
s
t
i
o
n
s
,
b
u
t
o
c
c
a
s
i
o
n
a
l
l
y
f
a
i
l
s
s
p
e
c
t
a
c
u
l
a
r
l
y
o
n
c
e
r
t
a
i
n
q
u
e
s
t
i
o
n
s
,
“
h
a
l
l
u
c
i
n
a
t
i
n
g
”
m
a
d
e
-
u
p
a
n
s
w
e
r
s
.
O
v
e
r
t
i
m
e
,
t
h
i
s
p
r
o
g
r
e
s
s
i
o
n
l
... | The a16z Investment Thesis on AI in Bio + Health _ Andreessen Horowitz |
16
parallelism rank 4, requiring 16 GPUs (two nodes) for one replica. To fully leverage the cluster’s
capabilities, we used 32-fold data parallelism. To optimize GPU utilization and reduce idle compute
bubbles, we maintained a micro-batch size of 1 and accumulated for 16 steps, resulting in a global
batch size of 512... | StarCoder_paper (1) |
3.1 SELF-DEBUGGING with Simple Feedback
The simplest form of automatic feedback is a sentence that just indicates the code correctness without
more detailed information. For instance, in text-to-SQL generation, the few-shot prompt provides the
feedback message “The SQL prediction above is correct!” for all correct SQL... | Teaching Large Language Models to Self-Debug |
fication and regression tasks; [66] which examines analogical reasoning in various text tasks; and [67] which
studies how LLMs can represent a rollout policy and world model in-context and then uses Q-learning to
drive policy improvement across a collection of toy environments with linguistic representations. Our use o... | LargeLanguageModelsasGeneralPatternMachines |
3.3 EMPIRICAL EVALUATION
We compare the proposed algorithm with competitive Flow-model-based (IDF by Hoogeboom et al.
(2019)) and VAE-based (BitSwap by Kingma et al. (2019)) neural compression algorithms using the
MNIST dataset. We first evaluate bitrates. As shown in Table 2, the PC (de)compressor achieved
compression ... | LOSSLESS COMPRESSION WITH PROBABILISTIC CIRCUITS |
1
INTRODUCTION | METAMATH |
Tokens per byte
(LT /LB)
0.2291
0.3103
0.2477
0.2434
0.3532
0.4412
0.2622
0.3436
0.2116
0.2183
0.2677
0.2765
0.2373
0.8137
0.3651
0.2430
0.3879
0.2627
0.4349
0.2688
0.1987
0.3103
Table 7: Tokens per byte for Pile components
TripAdvisor, SimplyHired, Associated Press, Post-
Media, The FCC etc. PHP error messages and
p... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Romanticism and Magic Realism are predicted as the most
aesthetically pleasing categories (0.63), while Minimalism
is the lowest ranked style with an average score of 0.49.
However, the mean aesthetic scores are similar across dif-
ferent styles and it is difficult to differentiate styles based
on the aesthetic scores, ... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
[45] Yang Song, Conor Durkan, Iain Murray, and Stefano Ermon.
Maximum likelihood training of score-based diffusion mod-
els. In Neural Information Processing Systems, 2021. 3
[46] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab-
hishek Kumar, Stefano Ermon, and Ben Poole. Score-based
generative modeling throug... | DiffusionModelAlignmentUsing Direct Preference Optimization |
3 Moûsai: Efficient Long-Context Music
Generation from Text
Our model Moûsai contains a two-stage training
process. In Stage 1, we use diffusion magnitude-
autoencoding (DMAE), which compresses the au-
dio waveform 64x using a diffusion autoencoder.
In Stage 2, we use a latent text-to-audio diffusion
model, to genera... | Moûsai |
● Lack of detailed context: Many tasks in the real economy require extensive context
about a particular company, project, or code-base. Current frontier systems are
generically competent, but lack this specific context and cannot learn it from the
available data. This might be addressed by access to additional pri... | Capabilities and risks from frontier AI |
50
Y. Dubois, T. Hashimoto, S. Ermon, and P. Liang. Improving Self-Supervised Learning
by Characterizing Idealized Representations, Dec. 2022. URL http://arxiv.org/abs/
2209.06235. arXiv:2209.06235 [cs, stat]. 24
D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman. With a little help from my
friends: Nea... | A Cookbook of Self-Supervised Learning |
Mitigating Gender Bias There is much work cataloging
how language models reflect the biases encoded in their
training data. However, while some work has explored
finetuning’s effects on bias in language models (Gira et al.,
2022; Kirtane et al., 2022; Choenni et al., 2021), or the
relationship between the corpus statisti... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
In Europe, it appears that there is a similar story. Evidence so far suggests
that the reach of fake news sites was limited: Using analytics data from
comScore and CrowdTangle, Fletcher et al. (2018) found that their sample of
fake news sites in France and Italy had an average monthly reach of 3.5 percent.
(For compari... | Social_Media_and_Democracy |
[83] Michael Zollh¨ofer, Matthias Niessner, Shahram Izadi,
Christoph Rehmann, Christopher Zach, Matthew Fisher,
Chenglei Wu, Andrew Fitzgibbon, Charles Loop, Christian
Theobalt, et al. Real-time non-rigid reconstruction using an
RGB-D camera. ACM Trans. Graphics, 33(4):156, 2014. | DynIBaR-NeuralDynamicImage-BasedRendering |
7. Related work
7.1. Program synthesis
Program synthesis consists of automatically generating a program that satisfies a task specification.
Possible ways of expressing the task include natural language descriptions, a set of input/output
examples, or a series of constraints. As a research topic, program synthesis has a ... | alphacode |
Programme or Research Masters, at the discretion of UCL, where:
a) There is space for additional students on the UCL Programme concerned, and
b) UCL is satisfied that the student is at least as well qualified as students who were able
to satisfy the standard entrance requirements at initial entry, and
c) UCL... | UCL Academic Manual |
we have R3(a1, a3) if and only if a3 = g2(g1(a)) = g3(a).
VP: Suppose both τ1 and τ2 are VP. Then there are two corresponding sets V C1 ⊆ V 1 and V C2 ⊆ V 2 of critical variables
and two corresponding bijections g1 : A1 → A2 and g2 : A2 → A3, such that Definition 32 is satisfied for both τ1 and τ2.
By definition, V... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
SoundStreamw2v-BERTMuLanDecoderRVQEncoderIntermediate LayerAudio NetworkAudio EmbeddingText NetworkText Embedding“Rock song withdistorted guitar”Adversarial and Reconstruction LossMLM loss and Contrastive Loss<Audio, Text> Contrastive Loss MusicLM: Generating Music From Text
Figure 2. Left: During training we extract ... | MusicLM |
rally, we replace them after every Transformer block with an input agnostic vector. Thus, both the
embeddings and subsequent Transformer block activations are treated as trainable parameters. For
more on prefix-layer tuning, see Section 5.1.
In Table 15, we show the evaluation results of LoRA+PE and LoRA+PL on WikiSQL a... | LORA |
https://github.com/LAION-AI/
Open-Instruction-Generalist, 2023.
[33] B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt
tuning. arXiv preprint arXiv:2104.08691, 2021.
[34] X. L. Li and P. Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv
preprint arXiv... | QLORA |
While we have not explored potential downstream applications of the generative models described
in this work, we believe Phenaki can have a positive impact in a variety of creative settings. In
general, many of the samples from the model will not perfectly correspond to the input caption or
the user’s intent; however, ... | PHENAKI- VARIABLE LENGTH VIDEO GENERATION FROM OPEN DOMAIN TEXTUAL DESCRIPTIONS |
5.2 Fine-tuning LLM for RAG
Optimizing the generator within the RAG model is a critical
aspect of its architecture. The generator’s role is to take the
retrieved information and produce relevant text, forming the
final output of the model. The optimization of the generator
aims to ensure that the generated text is both... | RAG forLargeLanguageModels-ASurvey |
Multi-head self-attention layer employs the self-attention
function with h heads in parallel. For an input sequence
X ∈ Rn×d with the sentence length n and hidden dimension
size of d. The query (Q), key (K), and value (V) vectors are
the transformation of input sequence X,
K = XWk + bk, Q = XWq + bq, V = XWv + bv,
(1... | Parameter-EfficientFine-TuningMethods |
[76] Zheng, L., Li, Z., Zhang, H., Zhuang, Y., Chen, Z., Huang, Y., Wang, Y., Xu, Y.,
Zhuo, D., Xing, E.P., et al.: Alpa: Automating inter-and {Intra-Operator} paral-
lelism for distributed deep learning. In: 16th USENIX Symposium on Operating
Systems Design and Implementation (OSDI 22), pp. 559–578 (2022)
[77] Micike... | Beyond Efficiency |
[435] Xu, N., S. Masling, M. Du, et al. Grounding open-domain instructions to automate web
support tasks. In K. Toutanova, A. Rumshisky, L. Zettlemoyer, D. Hakkani-Tür, I. Beltagy,
S. Bethard, R. Cotterell, T. Chakraborty, Y. Zhou, eds., Proceedings of the 2021 Conference
of the North American Chapter of the Associatio... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
as incorrect as the belief that a lightwave can only travel through space by causing
disturbances in the luminiferous aether.... [with] scientists ... misled by compelling but
incorrect analogies to the only systems they knew that had the required properties.
Ideas like database-style records for individuals, too... | The Next Decade in AI- |
e
s
t
h
a
t
g
e
n
e
r
a
t
e
u
n
i
q
u
e
d
a
t
a
f
r
o
m
p
a
r
t
i
c
i
p
a
n
t
s
| Product-Led AI _ Greylock |
Grand challenges
(Sec. 7)
Challenges
(1) Designing AGI benchmarks (2) Complete behavioral evaluation (3) Robustness evaluation (4) Dynamic and evolving evaluation
(5) Principled and trustworthy evaluation (6) Unified evaluation that supports all LLMs tasks (7) Beyond evaluation: LLMs enhancement
More challenging tas... | ASurveyonEvaluationofLargeLanguageModels |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.