text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
behavior as before: the outliers are that 1) quantization appears “easier” across the whole spectrum
of models at 4-bit, where even RTN performs relatively well, and 2) at 3-bit, RTN breaks down,
while GPTQ still provides good accuracy. We provide additional results in Appendix A.4. | GPTQ |
LOGGED MODELS AND
ML EXPERIMENTATION
During the experimentation phase of ML, data scientists
develop models designed to solve given tasks. After training
the models, they test them to evaluate their accuracy,
precision, recall (the percentage of correctly predicted
positive instances out of all actual positive in... | databrick 2023 report |
MODEL ANSWER (CORRECT BUT INVERTED SEMANTICS): 5 pies were baked and cut into 8 pieces each.
The 5 pies were then served to the guests. 8 x 5 = 40 pieces of pie. The guests ate 14 pieces of pie. 40 - 14 = 26
pieces of pie were left. The answer is 26. (cid:88) | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
3.1 Data Quantity
The explorations of the relationship between scal-
ing instruction data quantity and fine-tuned model
performance diverge in two directions. One branch
of research focuses on scaling down the instruction
data quantity to improve training efficiency. For ex-
ample, LIMA (Zhou et al., 2023a) carefully c... | DataManagementForLargeLanguageModels-ASurvey |
[45] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan,
Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe,
and Laurens Van Der Maaten. Exploring the limits of weakly
supervised pretraining. In ECCV, 2018. 1
[46] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan
Laptev, Josef Sivic, and Andrew Zisserman. End... | IMAGEBIND- One Embedding Space To Bind Them A |
Response: Hydrogen comes first as element number one.
Helium is second for balloons to have fun!
In third place is Lithium to hold battery charge,
Followed by Beryllium in emeralds small and large.
Boron’s number five to help us keep things clean.
Carbon’s next at six, and it’s found in every gene.
Nitrogen is seven, f... | Llama2 |
The main advantage of program-based tools is that they provide greater flexibility and customizability than
GUI-based tools, and users can build more sophisticated solutions for current problems. As a result, such tools
also have a steeper learning curve than GUI-based tools, they require a greater degree of technical e... | Tool Learning with Foundation Models |
arXiv:2304.03277, 2023.
[46] A. Poliak, J. Naradowsky, A. Haldar, R. Rudinger, and B. Van Durme. Hypothesis only baselines
in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and
Computational Semantics, pages 180–191, 2018.
[47] R. Pope, S. Douglas, A. Chowdhery, J. Devlin, J. Br... | QLORA |
Model Train CF Eval CF Aux Loss Percent Tokens Dropped
Sparse
Sparse
Sparse
Sparse
10.6%
0.3%
0.0%
0.0%
0.75
1.25
2.0
4.0
Yes
Yes
Yes
Yes
2.0
2.0
3.0
5.0
Sparse
Sparse
Sparse
Sparse
0.75
1.25
2.0
4.0
2.0
2.0
3.0
5.0
No
No
No
No
15.6%
2.9%
0.4%
0.0%
SuperGLUE (↑)
86.5 ± 0.21
86.7
85.8
86.4
85.7
85.8
85.9
86... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Classical Planning Automated planning (or classical planning) techniques can be used for com-
puting a sequence of actions that achieves a given goal [12, 13, 14]. Automated planning algorithms
have been widely used in robot systems. Shakey is the first robot that was equipped with a plan-
ning component, which was cons... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
3 AREN’T EXISTING SOLUTIONS GOOD ENOUGH?
The problem we set out to tackle is by no means new. Since the inception of transfer learning, dozens
of works have sought to make model adaptation more parameter- and compute-efficient. See Sec-
tion 6 for a survey of some of the well-known works. Using language modeling as an ... | LORA |
29
THE NEXT DECADE IN AI / GARY MARCUS
Ernest Davis, Noah Frazier-Logue and I, proposed a framework (Davis, Marcus, &
Frazier-Logue, 2017) that could help with this sort of challenge: a large set of
independently-motivated logical axioms—none specific to tea kettles, all of general
utility, largely co... | The Next Decade in AI- |
4.3.3 On the Performance of UniLM and SCLM
On the encoder-decoder setup, both the UniLM and SCLM objective performs better than the standard span
corruption objective in terms of aggregated and normalized overall gain. This shows that, in general, mixing
pre-training objectives is helpful. On the decoder setup, there i... | UL2- Unifying Language Learning Paradigms |
be emergent: while there is a performance over
the random baseline on a few of the non-emergent
tasks (e.g., english proverbs), these are not consid-
ered truly emergent, as this increased performance
is predictable based on the performance of smaller
models. This outcome, which aligns with previ-
ous results, serves a... | AreEmergentAbilitiesinLarge Language Models just In-Context |
1.2 Emergent Abilities vs Prompting
Techniques
The scaling up of LLMs facilitates the acquisition
of diverse competencies, which can be generally
grouped into two categories: The first group en-
compasses abilities, already described. The sec-
ond group encompasses various techniques, which
LLMs can benefit from, but... | AreEmergentAbilitiesinLarge Language Models just In-Context |
reduced other weight compression techniques must be used
to attain further savings. Our bottleneck adapters can be
much smaller, and still perform well.
Concurrent work explores similar ideas for BERT (Stickland
& Murray, 2019). The authors introduce Projected Atten-
tion Layers (PALs), small layers with a similar role... | Parameter-Efficient Transfer Learning for NLP |
In text-based games, all environment elements, such as locations, objects, characters, and actions,
are exclusively portrayed through textual descriptions. Agents utilize text commands to execute
manipulations like moving or tool use [432; 512; 514; 515]. Additionally, agents can convey emotions
and feelings through te... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
(Razumovskaia et al., 2023) introduce BEINFO, a
simple yet effective method that applies behavioral
tuning to aid information-seeking dialogue.
In
this work, the authors propose BEINFO, a simple
yet effective method that applies ‘behavioral
finetuning’ to increase the faithfulness of the
generated responses
information... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
Additionally, there are several long-standing chal-
lenges in the area of music generation: (1) music
generation at length, as most text-to-audio systems
(Forsgren and Martiros, 2022; Kreuk et al., 2022)
can only generate a few seconds of audio; (2) model
efficiency, as many need to run on GPUs for hours
to generate ju... | Moûsai |
We define LaMDA to be the model that incorporates all of the fine-tunings described above. We present their results in
Figure 4, and compare them to pre-training alone.
The figure shows that fine-tuning (in particular LaMDA) produces a significant improvement in quality, safety and
groundedness across all model sizes. Moreo... | LaMDA- Language Models for Dialog Applications |
address
tactics
used
the
by
to
42 The Honest Ads Act, S. 1989, 115th Cong. (2017).
44 See, e.g., Mina (2017), which discusses one initiative to develop “credibility indicators.”
43 Id. at §8.
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
268
Tim Hwang
algorithmic outpu... | Social_Media_and_Democracy |
on extremely large language models more accessible, for better or for worse. We believe that, in
time, such tools will become much easier to use and deploy, making the need to understand their
power and limitations even more stringent. | GPTQ |
ClockTEMUObservation:The screenshot shows a photo editing app on a smartphone with an image displayed on the screen...Thought: To complete this task, I should adjust the noise reduction levelusing the slider. Action:Action: swipe(23, "right", "medium")Observation:The image shows a screenshot of an alarm application on ... | AppAgents |
However, misinformation flags did not affect all partici-
pants equally. People who reported spending more time on
social media showed more resistance to both flags, suggest-
ing that perhaps these participants who spend more time
on social media have greater trust in online information or | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Figure 13: Prompts used to evaluate Code Llama on APPS.
34
H Addition results on responsible AI and safety
In this section, we present results of both pretrained and aligned LLMs on the three automatic safety
benchmarks from the perspectives of truthfulness, toxicity, and bias. The descriptions of the benchmarks ar... | CodeLlama2 |
language models. This paper outlines the develop-
ment paradigms of RAG in the era of LLMs, sum-
marizing three paradigms: Naive RAG, Advanced
RAG, and Modular RAG. It then provides a sum-
mary and organization of the three main compo-
nents of RAG: retriever, generator, and augmenta-
tion methods, along with key techn... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
Instead of performing single attention in each transformer block, multiple attentions in lower-
dimensional space have been observed to work better [554]. This observation gave rise to Multi-
Head Attention: For ℎ heads and dimension of tokens in the model 𝑑𝑚, the 𝑑𝑚-dimensional
query, key, and values are projected... | AReviewofDeepLearningTechniquesforSpeechProcessing |
tCO2eq = MWh × 0.385.
We apply the same formula to OPT and BLOOM
for fair comparison. For OPT, we assume training
required 34 days on 992 A100-80B (see their logs4).
Finally, we estimate that we used 2048 A100-80GB
for a period of approximately 5 months to develop
our models. This means that developing these mod-
els ... | LLaMA- Open and Efficient Foundation Language Models |
10
Model
Completion
Prompt
1M
8 layers
2.5M
8 layers
8.3M
8 layers
28M
8 layers
28M
8 layers
temper-
ature
0.8
21M
1 layer
33M
2 layers
Once upon a time there was a pumpkin. It was a very special pumpkin, it could speak. It was sad because
it couldn’t move. Every day, it would say
hello to the pumpkin.
On... | TinyStories-HowSmallCanLanguageModelsBeandStillSpeak CoherentEnglish? |
[Zhu et al., 2022] Wanrong Zhu, An Yan, Yujie Lu, Wenda
Xu, Xin Eric Wang, Miguel Eckstein, and William Yang
Wang.
Imagination-
arXiv preprint
guided open-ended text generation.
arXiv:2210.03765, 2022.
Visualize before you write:
[Zhu et al., 2023] Yutao Zhu, Huaying Yuan, Shuting Wang,
Jiongnan Liu, Wenhan Liu, Chen... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
arXiv:1901.04085, 2019. URL https://arxiv.org/abs/1901.04085.
[45] Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier,
and Michael Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings
of the 2019 Conference of the North American Chapter of the Associati... | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
Foundational work on neurosymbolic models is (D’Avila Garcez, Lamb, & Gabbay,
2009) which examined the mappings between symbolic systems and neural networks,
and showed important limits on the kinds of knowledge that could be represented in
conventional neural networks, and demonstrated the value in constructing mix... | The Next Decade in AI- |
What is the issue with my plant? Whatshould I do now?The image shows a leaf with brownspots on it. The spots are likelycaused by a fungal infection, whichcan be treated with a fungicide. It'simportant to identify the type offungus causing the spots in order tochoose the appropriate treatment.To treat the plant, you can... | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
4.2 Data Filtering
Data filtering is pivotal in directing training focus towards more informative samples, thereby eliminating irregular characters
or patterns, rather than concentrating on examples with lesser informational value.
Deduplication. A prime data filter is removing duplications, i.e., deduplication. This ... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
learning techniques, student will develop a 3D understanding of the urban crimes in a dynamic and
forecasting way, and contribute to the tradition literature on spatial analysis from an innovated angle by
adding the dynamic temporal and layers’ dimensions. | informatics-phd-projects-2022-23 |
RAG, introduced by Lewis et al. [Lewis et al., 2020] in
mid-2020, stands as a paradigm within the realm of LLMs,
enhancing generative tasks. Specifically, RAG involves an
initial retrieval step where the LLMs query an external data
source to obtain relevant information before proceeding to an-
swer questions or generat... | RAG forLargeLanguageModels-ASurvey |
discuss several key topics and open problems within the field. A repository for the
related papers at https://github.com/WooooDyy/LLM-Agent-Paper-List. | TheRiseandPotentialofLargeLanguageModel BasedAgents |
This study significantly contributes to the advancement of MT research and highlights the potential
of LLMs in enhancing translation capabilities. In summary, while LLMs perform satisfactorily
in several translation tasks, there is still room for improvement, e.g., enhancing the translation
capability from English to n... | ASurveyonEvaluationofLargeLanguageModels |
Dara, R. (2011). Intermediary Liability in India: Chilling Effects on Free Expression
on the Internet. https://cis-india.org/internet-governance/intermediary-liability-
in-india.pdf
Duguay, S., Burgess, J., & Suzor, N. (2018). Queer women’s experiences of patchwork
platform governance on Tinder, Instagram, and Vine. C... | Social_Media_and_Democracy |
and Effects in Engineering and Research, June 2022.
[95] PAI, “Data enrichment sourcing guidelines,” November 2022 2022. accessed 2023-03-13.
[96] PAI, “Responsible sourcing of data enrichment services,” June 2021 2021. accessed 2023-03-13.
[97] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal ... | gpt-4-system-card |
after viewing the flags. Dhar et al. (2016) proposed a rumor
control model where an “authenticated news agency” can
flood a social network with counter statements that dilute
the effects of misinformation (p. 56). Our study shows the
limitations of counter statements in practice when individual
users pick and choo... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
Act as OscarWildeGive shortanswersWhich is thebest city tolive in?London, ofcourse! Why is that? My dearfellow, Londonis the epitome ofsophistication and culture. It isthe city ofShakespeare and Dickens, ofthe great universities, ofthe museumsand galleries. It isthe city ofthe past,present, andfuture, wherethe old andt... | Llama2 |
[104] Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv preprint
arXiv:1606.08415 (2016)
[105] Bapna, A., Firat, O.: Simple, scalable adaptation for neural machine transla-
tion. In: Proceedings of the 2019 Conference on Empirical Methods in Natural
Language Processing and the 9th International Jo... | Beyond Efficiency |
5 Speech Processing Tasks
In recent times, the field of speech processing has gained significant attention due to its rapid
evolution and its crucial role in modern technological applications. This field involves the use of
diverse techniques and algorithms to analyse and understand spoken language, ranging from basic
... | AReviewofDeepLearningTechniquesforSpeechProcessing |
[195] Wei-Ning Hsu, Yu Zhang, Ron J Weiss, Yu-An Chung, Yuxuan Wang, Yonghui Wu, and James Glass. 2019. Disen-
tangling correlated speaker and noise for speech synthesis via data augmentation and adversarial factorization. In
ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICA... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Current methods to optimize the retrieval process mainly
include iterative retrieval and adaptive retrieval. These allow
the model to iterate multiple times during the retrieval process
or adaptively adjust the retrieval process to better accommo-
date different tasks and scenarios.
Iterative Retrieval
Regularly collec... | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
Such machinery is overwhelmingly powerful. All the world's web browsers, all the
world's operating systems, all the world's apps, and so forth are built upon them. (The
same tools are also, ironically, used in the specification and execution of virtually all of
the world's neural networks).
§
Yet historically m... | The Next Decade in AI- |
• A pre-trained model can be shared and used to build many small LoRA modules for dif-
ferent tasks. We can freeze the shared model and efficiently switch tasks by replacing the
matrices A and B in Figure 1, reducing the storage requirement and task-switching over-
head significantly.
• LoRA makes training more efficient... | LORA |
To evaluate toxicity in responses generated from our model, we use RealToxicityPrompts (Gehman
et al., 2020), a collection of sentence-level prompts that often elicit undesirable responses from
language models. We generate responses to 10K examples from RealToxicityPrompts using Star-
CoderBase with a minimum length of... | StarCoder_paper (1) |
0.05
0.51
0.22
0.04
0.17
0.02
0.09
0.44
0.10
0.16
0.11
0.10
0.05
0.51
0.22
0.04
0.17
0.02
Iterated DoReMi achieves performance of downstream-tuned weights on the GLaM dataset.
We employ iterated DoReMi on the GLaM dataset over 3 rounds. We find that the second and
third round domain weights are almost identical (Tabl... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
susceptibility to
misinformation. Need for closure refers to “the expedient desire for any firm
belief on a given topic, as opposed to confusion and uncertainty” (Jost et al.
2003, p. 348, italics in original). This motivation fosters two main behavioral
inclinations: the propensity to seize on readily available informa... | Social_Media_and_Democracy |
One surprising result is that we see the lowest performance on Hindi to English. This may arise from differences in
using general-purpose automated evaluation methods and using more precisely targeted examples and scoring methods
that target specific potential harms. The multi-sentence passages in the misgendering evalu... | PaLM 2 Technical Report |
We supplement the sandbox development framework with a
server that makes the sandbox information available to generative
agents and enables generative agents to move and influence the
sandbox environment. The server maintains a JSON data structure
that contains information about each agent in the sandbox world,
includi... | Generative Agents- Interactive Simulacra of Human Behavior |
Despite theory and expectation of large media effects,15, 16 media effect studies have found small to moderate effect
sizes.16, 17 Attenuated effect sizes have been attributed to (a) media content not being incorporated, and (b) media exposure
being only loosely measured. For example, a typical media impact study might... | Language models trained on media diets can predict public opinion |
Summary
Koala
Koala model page
Release date: April 2023
Koala is a LLaMA 7B and 13B models fine-tuned with publicly available dialog data by an
academic team at UC Berkeley.
Training
The training data includes filtered data from multiple datasets.
ShareGPT – 30k
Human ChatGPT Comparison Corpus – 87k
Open Instructi... | A brief history of LLaMA models - AGI Sphere |
Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Anselm Levskaya,
Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference, 2022.
Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides,
Sarah H... | Llama2 |
of the Tanks and Temples dataset [15], including large-scale
indoor/outdoor scenes. Each scene contains 263 to 1107
images captured using a hand-held monocular RGB camera.
The ground truth is obtained using a LiDAR sensor.
Implementation details. Our hash encoding resolution
spans 25 to 211 with 16 levels. Each hash en... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
Sadly, we are not out of the woods yet, though. Hybrid models that combine powerful
data-driven learning techniques with the representational and computational resources
of symbol-manipulation may be necessary for robust intelligence, but they are surely
not sufficient. In what follows I will describe three further... | The Next Decade in AI- |
enrichtheforecastexplanations[57].
When designing a system providing explanations for AI models,
it must be considered the explanations must serve multiple stake-
holders (and thus target different user profiles [58]), serve different
purposes, and that their effectiveness must be assessed quantitatively
(throughalgori... | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
3.4. Parameter/Performance trade-off
The adapter size controls the parameter efficiency, smaller
adapters introduce fewer parameters, at a possible cost to
performance. To explore this trade-off, we consider different
adapter sizes, and compare to two baselines: (i) Fine-tuning
of only the top k layers of BERTBASE. (ii... | Parameter-Efficient Transfer Learning for NLP |
https://a16z.com/2023/06/21/ai-bio-health-thesis/
8/9
G
e
n
e
r
a
t
i
v
e
A
I
:
T
h
e
N
e
x
t
C
o
n
s
u
m
e
r
P
l
a
t
f
o
r
m
b
y
C
o
n
n
i
e
C
h
a
n
a
n
d
J
u
s
t
i
n
e
M
o
o
r
e
T
h
e
G
e
n
e
r
a
t
i
v
e
A
I
R
e
v
o
l
u
t
i
o
n
i
n
G
a
m
e
s
b
y
J
a
m
e
s
G
w
e
r
t
z
m
a
n
a
n
d
... | The a16z Investment Thesis on AI in Bio + Health _ Andreessen Horowitz |
The increasing demand for ML solutions and the growing
availability of technologies have led to a significant
increase in experimentation and production, two distinct
parts of the ML model lifecycle. We look at the logging and
registering of models in MLflow, an open source platform
developed by Databricks, to und... | databrick 2023 report |
4.1
IMPLEMENTATION DETAILS
We implement everything in PyTorch (Paszke et al., 2017) and to limit our gains from the ”soft-
ware lottery” (Hooker, 2021) we do not use specialized implementations, which would further bias
results towards well-established components. We keep everything on the implementation level of
the... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
2https://ai.google/principles/
5
4.2 Role-specific metrics: Helpfulness and Role consistency
The foundation metrics (quality, safety, and groundedness) measure attributes that we find important for dialog agents
in general. However, they are not dependent on any application-specific role that an agent may be designed ... | LaMDA- Language Models for Dialog Applications |
task parsing and planning. By injecting several demonstrations into the prompts, HuggingGPT allows
the large language model to better understand the intention and criteria for task planning. Each
demonstration is a group of input and output on task planning - the user’s request and the expected
task sequence to be pars... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
training, a KL reward coefficient of λKL = 0.001 (4.1), PPO clipping (cid:15) = 0.2, discount factor γ = 1, and
no entropy bonus. Furthermore, in PPO, we re-iterate over the same sample K times (see Algorithm 1 in
[Schulman et al., 2017]), with higher K typically leading to more stable results. We used K = 1 for the
RLH... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
The paper is organized as follows. In Sec. 2, we provide the basic information of LLMs and AI
model evaluation. Then, Sec. 3 reviews existing work from the aspects of “what to evaluate”. After
that, Sec. 4 is the “where to evaluate” part, which summarizes existing datasets and benchmarks.
Sec. 5 discusses how to perfor... | ASurveyonEvaluationofLargeLanguageModels |
[12] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao,
Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping
Wang. Tore: Token reduction for efficient human mesh re-
covery with transformer. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 15143–
15155, 2023. 2
[13] Laura Downs, Anthony Fra... | Wonder3D |
3. Method
Our goal is to learn a single joint embedding space for all
modalities by using images to bind them together. We align
each modality’s embedding to image embeddings, such as
text to image using web data and IMU to video using video
data captured from egocentric cameras with IMU. We show
that the resulting em... | IMAGEBIND- One Embedding Space To Bind Them A |
4 Application and Experiment
In this section, we aim to explore the applications of tool learning and investigate the efficacy and limitations
of state-of-the-art foundation models in utilizing tools. We select 17 representative tools for evaluation and
place the main results in this section. For more case studies of C... | Tool Learning with Foundation Models |
Finally, Table 2 shows more detailed downstream task comparisons for large publicly-available models,
grouped into comparable sizes. We bold the results that are the best for each task and model size group.
Each model family has at least one model that is best for some tasks. In this table, we also include results for
... | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
Laurenc¸on, H., Saulnier, L., Wang, T., Akiki, C., del Moral,
A. V., Scao, T. L., Werra, L. V., Mou, C., Ponferrada,
E. G., Nguyen, H., Frohberg, J., ˇSaˇsko, M., Lhoest, Q.,
McMillan-Major, A., Dupont, G., Biderman, S., Rogers,
A., allal, L. B., Toni, F. D., Pistilli, G., Nguyen, O.,
Nikpoor, S., Masoud, M., Colombo, ... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
35
References
[1] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.
Advances in neural information processing systems, 33:1877–1901, 2020.
[2] Long Ouyang, Jeffrey Wu,... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
oshowonlythemountainswithanelevationabove8,500meters.FirstlyIneedtoconverttheelevationcolumntonumeric.Action:python_repl_astActionInput:convert_to_numeric(df,'Elevation(m)')Observation:Thought:NowIcanfilterthetable.Action:python_repl_astActionInput:filter_df(df,'SELECT*FROMdfWHERE"Elevation(m)">8500')Observation:Thought... | Tool Learning with Foundation Models |
(cid:2)∥fθ(xσt, σt) − vσt∥2
2
(cid:3) ,
Et∼[0,1],σt,xσt
(1)
= ασtϵϵϵ − βσtxxx0, for which
where vvvσt = ∂xxxσt
σt
we define ϕt := π
2 σt, and obtain its trigonometric
values ασt := cos(ϕt), and βσt := sin(ϕt).
3.1.2 DDIM Sampler for Denoising
The denoising step uses ODE samplers to turn noise
into a new data point... | MOUSAI |
Illustration generated with Midjourney
Wave 4: Killer apps emerge (Now) With the platform layer solidifying, models continuing to get better/faster/cheaper, and model access trending to free and open source, the application layer is ripe for an explosion of creativity.
Just as mobile unleashed new types of application... | Generative AI A Creative New World Sequoia Capital |
LLMs
evaluation
Other
applications
General
benchmarks
Specific
benchmarks
Multi-modal
benchmarks
Engineering: Bubeck et al. [13] / Liu et al. [116] / Pallagani et al. [140] / Sridhara et al. [171] / Valmeekam et al. [184]
Valmeekam et al. [183] / Zhuang et al. [250]
Medical queries: Chervenak et al. [19] / Duong ... | ASurveyonEvaluationofLargeLanguageModels |
specific groups or individuals (Castle 2012). By providing efficient ways to
reach new audiences and disseminate hateful language, the Internet enables
hate groups to be well represented in the digital realm, fostering a sense of
community among their members, and attracting the attention of journalists
and everyday citi... | Social_Media_and_Democracy |
3 Method
This section first uses a running example to investigate ways for formulating planning prompts in
PDDL and then introduces the LLM+P method.
3We
refer
the
readers
to
introtopddl2.pdf as a good introduction to PDDL.
https://www.cs.torontnaturalo.edu/~sheila/2542/s14/A1/
3
LLMPlanLLMProblem PDDLPlannerP... | LLM+P- Empowering Large Language Models with Optimal Planning Proficiency |
tool-interactive critiquing. CoRR, abs/2305.11738, 2023.
[163] Lewis, M., Y. Liu, N. Goyal, et al. BART: denoising sequence-to-sequence pre-training
for natural language generation, translation, and comprehension. In D. Jurafsky, J. Chai,
N. Schluter, J. R. Tetreault, eds., Proceedings of the 58th Annual Meeting of th... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
exp(0)+10·exp(−0.5) ≈ 0.142 to
exp(0)
exp(0)
9
Figure 3: Sparse models are prone to overfit. We plot train and validation curves for our ST-MoE-
L and a dense-L models fine-tuned on the CB task (250 train sequences) and ReCoRD (138k train
sequences). In both cases, the sparse model learns more quickly on the train p... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
s
e
m
a
n
t
i
c
n
e
u
r
o
n
s
w
h
i
c
h
c
o
r
r
e
s
p
o
n
d
t
o
m
u
l
t
i
p
l
e
s
e
m
a
n
t
i
c
c
o
n
c
e
p
t
s
.
W
h
i
l
e
o
u
r
e
x
p
l
a
n
a
t
i
o
n
t
e
c
h
n
i
q
u
e
c
a
n
a
n
d
o
f
t
e
n
d
o
e
s
g
e
n
e
r
a
t
e
e
x
p
l
a
n
a
t
i
o
n
s
a
l
o
n
g
t
h
e
l
i
n
e
s
o
f
"
X
a
... | Language models can explain neurons in language models |
Table 22: Few-shot exemplars for full chain of thought prompt for the last letter concatenation task.
PROMPT FOR LAST LETTER CONCATENATION
Q: Take the last letters of the words in "Elon Musk" and concatenate them.
A: The last letter of "Elon" is "n". The last letter of "Musk" is "k". Concatenating them is "nk". The an... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
[33] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis,
Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Mor-
gan McGuire, and Sanja Fidler. Neural geometric level of
detail: Real-time rendering with implicit 3d shapes. In Pro-
ceedings of the IEEE/CVF Conference on Computer Vision
and Pattern Recognition, p... | Neuralangelo- High-Fidelity Neural Surface Reconstruction |
We did not expect these changes to significantly impact any experimental findings in the paper, and we reran all analyses
and evaluations on the new models to confirm this was indeed the case. All experiments in the paper report results from
this updated version of the suite. We chose to rerun the training runs in order t... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
A Details, Analysis, and Evaluations of Supervised Training
A.1 Context Distillation
For context distillation, we follow the prescription from [Askell et al., 2021]. Specifically, we first generate
data in the following way:
1. We prepend the ‘HHH prompt’ (i.e., a set of prompts designed to elicit helpfulness, harmles... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
In this section we prove that in Example 1, action a3 and the bid profile b1 = (0, 0, 1+γ), b2 = (0, 1+
(cid:15), 0) and b(cid:96) = (0, 0, 0) ∀(cid:96) > 2 constitute an equilibrium. To do so, we show that the agent maximizes
his utility, and that each principal cannot benefit from deviating and switching her bid. Note ... | Incomplete Information VCG Contracts for Common Agency |
2.2 Costs
Running all of our experiments cost about $5000
in GPU costs. We gratefully acknowledge our
compute sponsor Paperspace for their generosity
in making GPT4All-J training possible. Between
GPT4All and GPT4All-J, we have spent about
$800 in OpenAI API credits so far to generate
the training samples that we open... | 2023_GPT4All-J_Technical_Report_2 |
GRAND TOTAL (28 ds.) 13.4M >1k
2.8M 13
10.8M >900 277
77
555
♢
♣
♡
parate skeleton annotation formats in these datasets, which
has rarely been addressed in the literature so far. (2) We pro-
pose affine-combining autoencoders (ACAE), a novel linear
dimensionality reduction technique applicable to keypoint-
based r... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
u
r
e
x
a
m
p
l
e
s
a
s
s
u
m
e
,
u
n
l
e
s
s
o
t
h
e
r
w
i
s
e
s
t
a
t
e
d
.
P
r
a
t
a
n
d
R
u
s
t
i
c
h
i
n
i
(
2
0
0
3
)
s
t
u
d
y
c
l
a
s
s
i
c
c
o
n
t
r
a
c
t
s
:
F
o
r
e
a
c
h
a
g
e
n
t
n
,
t
h
i
s
i
s
a
m
a
p
p
i
n
g
f
r
o
m
t
h
e
a
g
e
n
t
'
s
a
c
t
i
o
n
s
t
o
... | Principal-agent VCG contracts - ScienceDirect |
Dominik Stammbach. 2021. Evidence selection
as a token-level prediction task. In Proceed-
ings of the Fourth Workshop on Fact Extrac-
tion and VERification (FEVER), pages 14–20,
Dominican Republic. Association for Com-
putational Linguistics. https://doi.org
/10.18653/v1/2021.fever-1.2
Asher Stern, Roni Stern,
Ido Da... | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
Table 22: Hyperparameters used for all finetuned models.
Another important hyperparameter for instruction finetuning is the sampling rates for each tasks. Within the
four mixtures (Muffin, T0-SF, NIV2, and CoT defined in Figure 2), we use the number of examples as the
weight of each task. We apply the maximum cap for each ... | Scaling Instruction-Finetuned Language Models |
4Asking models to identify the word displayed as ASCII art, https://github.com/google/BIG-bench/tree/main/bigbench/benchmark_tasks/ascii_word_recognition
5Asking models to choose the English sentence with adjectives in the "correct" order within two choices, https://github.com/google/BIG-
bench/tree/main/bigbench/bench... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
Recent advances in machine learning have led to an acceleration of interest in research on artificial intelligence (AI).
This fostered the exploration of possible applications of AI in various domains and also prompted critical discussions
addressing the lack of interpretability, the limits of machine intelligence, pote... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
[97] J. Frankle, G. K. Dziugaite, D. Roy, and M. Carbin, “Pruning neural
networks at initialization: Why are we missing the mark?” in Proc. Int.
Conf. Learn. Representations, 2021.
[98] D. C. Mocanu, E. Mocanu, P. Stone, P. H. Nguyen, M. Gibescu,
and A. Liotta, “Scalable training of artificial neural networks with
ada... | Parameter-EfficientFine-TuningMethods |
6.1. Robot Environments / Tasks
Our three robot environments (Fig. 1) include a Task and
Motion Planning (TAMP) domain where a robot has to
manipulate (grasp and stack) objects, a table-top pushing
environment, and a mobile manipulation domain. In each
domain, PaLM-E is trained on expert data from that do-
main. In man... | PaLM-E- An Embodied Multimodal Language Model |
Computing Machinery, New York, NY, USA, 230–242. https://doi.org/10.1145/3519391.3519401
[39] Shunichi Kasahara, Jun Nishida, and Pedro Lopes. 2019. Preemptive Action: Accelerating Human Reaction Using Electrical Muscle
Stimulation Without Compromising Agency. In Proceedings of the 2019 CHI Conference on Human Factors... | Society’sAttitudesTowardsHumanAugmentation |
Aside from key actors, then, other mechanisms of diffusion include a mix of
biases – cognitive, social, and algorithmic (Shao et al. 2017). Information
diffusion tends to be bounded by limited attention resources; information
disseminated during an “attention burst” – a period of demand for a given
topic – is more like... | Social_Media_and_Democracy |
We optimize Eq. 11 using the Adam optimizer [28] with
hyperparameters β1 = 0.9 and β2 = 0.99. We set the
learning rate to 5 × 10−4 for θc (the canonical MLP), and
5 × 10−5 for all the others. We use 128 samples per ray.
The optimization takes 400K iterations (about 72 hours) on
4 GeForce RTX 2080 Ti GPUs. We apply dela... | HumanNeRF- Free-viewpoint Rendering of Moving People from Monocular Video |
2
2
0
2
v
o
N
7
]
L
C
.
s
c
[
5
v
9
2
6
3
0
.
2
0
2
2
:
v
i
X
r
a
Survey of Hallucination in Natural Language Generation | SurveyofHallucinationinNatural Language Generation |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.