Datasets:

andy8
/

HMCF

License:

Dataset Preview

Split (1)

train

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

The dataset generation failed because of a cast error

Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 3 new columns ({'id', 'metadata', 'name'}) and 4 missing columns ({'annotation', 'text', 'label1', 'label0'}).

This happened while the json dataset builder was generating data using

zip://HMCF-Dataset/ICLR/ICLR_2017/ICLR_2017_paper/ICLR_2017_100_paper.json::/tmp/hf-datasets-cache/medium/datasets/79880807925923-config-parquet-and-info-andy8-HMCF-1946d9f9/hub/datasets--andy8--HMCF/snapshots/830e168a996430d969b60d1b69c444c49d6df319/HMCF-Dataset.zip, [/tmp/hf-datasets-cache/medium/datasets/79880807925923-config-parquet-and-info-andy8-HMCF-1946d9f9/hub/datasets--andy8--HMCF/snapshots/830e168a996430d969b60d1b69c444c49d6df319/Blind-Annotation-Record.zip (origin=hf://datasets/andy8/HMCF@830e168a996430d969b60d1b69c444c49d6df319/Blind-Annotation-Record.zip), /tmp/hf-datasets-cache/medium/datasets/79880807925923-config-parquet-and-info-andy8-HMCF-1946d9f9/hub/datasets--andy8--HMCF/snapshots/830e168a996430d969b60d1b69c444c49d6df319/HMCF-Dataset.zip (origin=hf://datasets/andy8/HMCF@830e168a996430d969b60d1b69c444c49d6df319/HMCF-Dataset.zip)]

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
                  writer.write_table(table)
                File "/usr/local/lib/python3.12/site-packages/datasets/arrow_writer.py", line 674, in write_table
                  pa_table = table_cast(pa_table, self._schema)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2272, in table_cast
                  return cast_table_to_schema(table, schema)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2218, in cast_table_to_schema
                  raise CastError(
              datasets.table.CastError: Couldn't cast
              name: string
              metadata: struct<source: string, title: string, authors: list<item: string>, emails: list<item: null>, section (... 365 chars omitted)
                child 0, source: string
                child 1, title: string
                child 2, authors: list<item: string>
                    child 0, item: string
                child 3, emails: list<item: null>
                    child 0, item: null
                child 4, sections: list<item: struct<heading: string, text: string>>
                    child 0, item: struct<heading: string, text: string>
                        child 0, heading: string
                        child 1, text: string
                child 5, references: list<item: struct<title: string, author: list<item: string>, venue: string, citeRegEx: string, short (... 32 chars omitted)
                    child 0, item: struct<title: string, author: list<item: string>, venue: string, citeRegEx: string, shortCiteRegEx:  (... 20 chars omitted)
                        child 0, title: string
                        child 1, author: list<item: string>
                            child 0, item: string
                        child 2, venue: string
                        child 3, citeRegEx: string
                        child 4, shortCiteRegEx: string
                        child 5, year: int64
                child 6, referenceMentions: list<item: struct<referenceID: int64, context: string, startOffset: int64, endOffset: int64>>
                    child 0, item: struct<referenceID: int64, context: string, startOffset: int64, endOffset: int64>
                        child 0, referenceID: int64
                        child 1, context: string
                        child 2, startOffset: int64
                        child 3, endOffset: int64
                child 7, year: int64
                child 8, abstractText: string
                child 9, creator: string
              id: string
              to
              {'text': Value('string'), 'label0': Value('string'), 'label1': Value('string'), 'annotation': Value('string')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1347, in compute_config_parquet_and_info_response
                  parquet_operations = convert_to_parquet(builder)
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 980, in convert_to_parquet
                  builder.download_and_prepare(
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 884, in download_and_prepare
                  self._download_and_prepare(
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 947, in _download_and_prepare
                  self._prepare_split(split_generator, **prepare_split_kwargs)
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1736, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1889, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 3 new columns ({'id', 'metadata', 'name'}) and 4 missing columns ({'annotation', 'text', 'label1', 'label0'}).
              
              This happened while the json dataset builder was generating data using
              
              zip://HMCF-Dataset/ICLR/ICLR_2017/ICLR_2017_paper/ICLR_2017_100_paper.json::/tmp/hf-datasets-cache/medium/datasets/79880807925923-config-parquet-and-info-andy8-HMCF-1946d9f9/hub/datasets--andy8--HMCF/snapshots/830e168a996430d969b60d1b69c444c49d6df319/HMCF-Dataset.zip, [/tmp/hf-datasets-cache/medium/datasets/79880807925923-config-parquet-and-info-andy8-HMCF-1946d9f9/hub/datasets--andy8--HMCF/snapshots/830e168a996430d969b60d1b69c444c49d6df319/Blind-Annotation-Record.zip (origin=hf://datasets/andy8/HMCF@830e168a996430d969b60d1b69c444c49d6df319/Blind-Annotation-Record.zip), /tmp/hf-datasets-cache/medium/datasets/79880807925923-config-parquet-and-info-andy8-HMCF-1946d9f9/hub/datasets--andy8--HMCF/snapshots/830e168a996430d969b60d1b69c444c49d6df319/HMCF-Dataset.zip (origin=hf://datasets/andy8/HMCF@830e168a996430d969b60d1b69c444c49d6df319/HMCF-Dataset.zip)]
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

text string	label0 string	label1 string	annotation string
It is prefer to provide detailed description or pseudocode for this step.	none	request	request
For the modeling contribution, although it shows some improvements on the benchmarks and some nice analysis, the paper really doesn’t explain well the intuition of this “write” operation/Scratchpad (also the improvement of Scratchpad vs coverage is relatively limited). Is this something tailored to question generation?...	evaluative	request	request
This leads to over-confident predictions which is problematic particularly in an active learning scenario.	structuring	evaluative	evaluative
One would expect the proposed approach to work better than diagonal preconditioning on a per-iteration basis (at least in terms of training loss).	fact	evaluative	evaluative
Simply because for continuous variables similar experiments have been reported before	evaluative	fact	fact
The paper used very restricted Gaussian distributions for the formulation.	fact	evaluative	fact
When a new image is presented, how is the probability mass distributed across previously seen classes versus across unseen ones?	none	request	request
At the very least we need another "partial" sign in front of the "\delta" function in the numerator.	request	evaluative	request
The algorithm here seems to build off of the former algorithm; essentially replacing a single hard thresholding step with an IHT-like step.	fact	structuring	fact
While most related work was covered well, I believe the authors could have a more up-to-date list of recent work that reconstructs triangle-mesh representations from images [A-C] (especially since several of these methods has an architecture that involves encoding and subsequent compositional refinement).	request	evaluative	request
This work is relevant to researchers in the field of continual/life-long learning, since it proposes a framework, which should be possible to integrate into different approaches in this field.	structuring	evaluative	evaluative
The ostensible goal of learning more about observation representations is mostly preliminary — and this direction holds promise of for a stronger set of findings.	evaluative	fact	evaluative
I am concerned though why the authors didn’t compare to adaptive optimizers such as ADAM and ADAGRAD and how the performance compares with population based training techniques.	evaluative	request	request
I don't think this is really a fair comparison; I would have liked to have seen results for the unmodified reward function.	request	evaluative	request
This restriction exists in Dreamer, and the method cannot be applied to discrete control tasks unless approximation techniques such as Gumbel-softmax are used.	evaluative	fact	fact
The reason for this is only given in a single sentence at the end of Section 6, so it is a little confusing.	request	evaluative	evaluative
- Minor, but some of these citations can be updated.	request	structuring	request
It is also not clear why Table. 3 does not report the Bayes baseline results.	evaluative	request	request
1. Missing comparison with parameter counting bounds: there has been a long line of research on generalization bounds for RNNs by obtaining bounds on the VC dimension of the function class [1, 2] which provide generalization bounds for various non-linearities.	request	evaluative	request
The dataset is designed carefully so that it is very unlikely there will be any duplicate between train/test split and the difficulty can be controlled.	structuring	evaluative	evaluative
Experiments are conducted on a simple toy data set, as a proof of concept, and on data from ModelNet10 and ModelNet40.	none	structuring	structuring
Similarly, you did not indicate what the deterministic version of your model is.	request	evaluative	request
That is, authors assumed that they have a degradation function F and all the inference process is just based on this known function.	fact	evaluative	fact
I would suggest comparing with CW attack under different sets of hyper-parameters.	evaluative	request	request
However, there is a lot of important material in the Appendix, which I think may be relevant to the readers.	evaluative	fact	evaluative
The paper is well written in general, the experiments are extensive.	structuring	evaluative	evaluative
- In Section 4.2, in the second paragraph, you refer to Appendix F and describe “sharp upward jump at collapse” in D’s loss.	request	fact	fact
In addition, HSIC is a non-adaptive test, but your test is adaptive, so a fairer comparison would be to a modern adaptive test such as "An Adaptive Test of Independence with Analytic Kernel Embeddings."	none	request	request
Thus the main novelty claim of the paper needs to be hedged appropriately.	evaluative	request	evaluative
However, these were recently NeurIPS papers, and the text is not yet out, so I don't think this should affect the authors' independent work (and also the product part is new).	evaluative	fact	evaluative
In fact, assuming that such methods outperform general-purpose models, we could investigate why and where this is the case (in fact the proposed dataset is very useful for this).	request	fact	fact
The experimental validation of the proposed approach can also be further improved, see more specific comments below.	fact	request	request
The input 3D points are sampled from a unit sphere.	fact	structuring	fact
Additionally, authors combined W_U with W_L with a mixture 20:1, i.e., the s in Eqs(6, 13, 14) is smaller than 0.05.	fact	structuring	structuring
As a result, the main question in evaluating this paper is on the significance of the result and the generality of the “sufficiently bilinear” condition.	structuring	evaluative	structuring
This is a non-differentiable process and relies on maintaining a large pool of candidates out of which best ones are chosen with the highest fitness.	fact	structuring	structuring
I believe that a more challenging experiment should be conducted e.g. using celebA dataset.	evaluative	request	request
The question asked is whether or not pretrained mean-field RBMs can help in preventing adversarial attacks.	structuring	request	structuring
In terms of the empirical results, the better performance of DEBAL compared to a single MC-Dropout model is not supervising as Beluch et al. (2018) already demonstrated that an ensemble is better than a single MC-Dropout.	fact	evaluative	fact
1. The method used a latent dynamics model, which avoids reconstruction of the future images during inference.	evaluative	fact	fact
The number of parameters C is smaller than the number of free parameters in the image X, so this results in a predictive model that can be used for compression, denoising, inpainting, superresolution and other inverse problems.	structuring	fact	fact
It is unclear whether the data augmentation techniques is applied only at training time or also at test time.	evaluative	request	evaluative
This average is linked to the notion of an eligibility trace, and ties into some recent biological work that shows the role of dopamine in retroactively modulating synaptic plasticity.	structuring	fact	structuring
In practice however, it appears from the experimental section that the domain mapping is learned offline, and then frozen for the meta-learning phase.	structuring	fact	structuring
For example, how do I even know that the oracle in question exists? What are the particular assumptions under which it exists? What are the requirements on the training data, optimization ability, generalization error, etc. How do we know that we can create in practice ML learning models that are sufficiently accurate ...	none	request	request
I think that the other reviewers make a number of valid points, especially with regards to the theoretical analysis of the paper.	evaluative	social	evaluative
it would like that the authors provide more intuition why these improvements occur and also outline the limitations of their approach.	none	request	request
In the longer run it would be extremely beneficial to the community if this approach is applied to the standard benchmarks as set out in [2].	request	fact	fact
As other works, the solution is based on a proper initialization of the dictionary.	structuring	fact	fact
One of the main contributions of this work is that the proposed analysis is focusing on last iterate convergence guarantees for the HGD.	structuring	evaluative	evaluative
This work extends Schlichtkrull et al. (2018) by adding attention in two distinct ways: attention between pairs of nodes per relation, and attention between pairs of nodes averaged over all relations.	structuring	evaluative	evaluative
Extensive experiments with varying losses, architectures, hyperparameter settings are conducted to show self-modulation improves baseline GAN performance.	evaluative	fact	fact
I do not have major comments about the paper itself, although I did not check the technical details super carefully.	fact	evaluative	evaluative
The idea that multiple mappings will produce better results than a single mapping is reasonable given previous results on ensemble methods.	evaluative	structuring	evaluative
Their experiments confirm all the hypotheses (DSO-NAS can find architectures, having small FLOP counts, having good performances on CIFAR-10 and ImageNet).	structuring	fact	fact
In particular, evaluation for the classification task should be compatible with the proposed model, which would give a much better picture of the learned representations.	request	fact	request
Given the authors' rebuttal to all reviews, I am upgrading my score to a 6.	fact	evaluative	fact
Thus the proof is concluded by the linear (resp. sublinear) convergence of gradient descent (resp. stochastic GD) under PL assumption.	fact	structuring	structuring
This is unlike existing methods which use model-free or planning methods on simulated trajectories to learn the optimal policy.	fact	structuring	structuring
In the end, no solution that can ensure quality and stability is found, except having prohibitively large amounts of data (~300M images).	structuring	evaluative	evaluative
The author never explains. E.g., link to NRMSE and PFC to the Table.	request	evaluative	request
The MNIST is a relatively simple experiment, and I would like to see how the method works in more challenging problems.	request	evaluative	request
Consider $x$ a binary vector and reward equal to the parity $S(x) = \sum{x_j} % 2$.	structuring	other	structuring
It would probably help to position the VAE component more precisely w.r.t. one of the two baselines, by indicating the differences.	request	evaluative	request
In examining reward curves (generally extrinsic during testing), ‘curiosity-based’ reward generally works with the representation effectiveness varying across different testbeds.	structuring	fact	structuring
The most interesting one is the prediction of its dimensions by the CSLB features, which reveals a nice clustering in the different SPoSE dimensions.	structuring	evaluative	evaluative
The message of synthetic experiments would be stronger if more of them were available and if the comparison between LOE, TSTE, and OENN was made on more of them.	evaluative	request	request
Experimental results of different GNN architectures w/o different PT for different tasks are provided.	structuring	evaluative	structuring
It is known that SGD with fixed step-size can not find the optimal for convex (perhaps, also simple) problems.	fact	request	fact
And the results are obtained by G(z).	structuring	fact	structuring
In addition to the extensive experimentation on different settings showing performance improvements, the authors also present an ablation study, that shows the impact of the method when applied to different layers.	structuring	fact	structuring
As indicated in the recent literature, enforcing shift-invariance does help to improve the performance of a CNN on classification accuracy and the robustness with respect to image shift.	fact	evaluative	evaluative
- Although the paper introduces the generalization bound for MDL, it does not give new formulation or algorithm to handle MDL (MULANN handles only the class asymmetry when domains involve distinct sets of classes and it has nothing to do with MDL).	evaluative	fact	evaluative
There are a few grammatical/spelling errors that need ironing out.	request	evaluative	request
Additionally, CNNs are not a particularly good architecture for fMRI, as fMRI is not locally translation invariant (see Aydöre ICML 2019 for instance).	fact	evaluative	fact
- Regarding the diversity/fidelity tradeoff using different truncation thresholds, I think constraining the norm of the sampled noise vectors to the exact threshold value (by projecting the samples on the 0-centered hyper-sphere of radius = threshold) could yield even more interesting or more informative Figures, as ob...	request	fact	request
The authors pose an interesting hypothesis, but it would gain a lot of credibility if they could provide an empirical analysis of an algorithm that uses ME reasoning to improve learning in a realistic setting or if they could at least perform some quantitative analysis of the effects of ME bias on task performance for ...	evaluative	request	request
Because these differences aren't explained, the synthetic tasks in the experimental section make this approach look artificially good in comparison to Hartford et al.	evaluative	request	evaluative
But first, it’s not obvious why this would be a good thing (or a bad thing for that matter)	request	evaluative	evaluative
Experiments show that the method substantially outperforms sound typing in the TypeScript compiler, as well as a recent method based on deep neural networks.	structuring	fact	structuring
It would be nice to see a more direct comparison between the three definitions of spherical convolution (general SO3, isotropic S2, and anisotropic S2).	evaluative	request	request
Similar curves could also be produced with the hyper-sphere projection proposed above to have a slightly clearer idea of the behavior on the limit of that hyper-sphere.	evaluative	request	evaluative
The theoretical proof depends on the convexity assumption, I would also suggest comparing the proposed attack with CW and other benchmarks on some simple models that satisfy the assumptions.	evaluative	request	request
The paper is presented in a clear manner, with the objectives and analysis techniques delineated in the main paper.	structuring	evaluative	evaluative
- For the cart-pole task, the paper states that the reward is modified "to exclude any cost objective".	fact	structuring	structuring
It makes sense, and I expect the proposed model may benefit from its design for long-range spatio-temporal feature learning.	evaluative	fact	evaluative
Using spectral regularization to improve robustness is not new, but it's interesting to combine spectral regularization and adversarial training.	evaluative	structuring	evaluative
The problem is very interesting.	structuring	evaluative	evaluative
The introduction contains too much related work, which should be divided in another section.	request	evaluative	request
this makes a lot of sense especially given that CPC / wav2vec recovers phonemes and quantizing the phonemes will recover a language-like version of the raw audio. And running BERT across those tokens will allow you to capture the dependencies at the phoneme level.	structuring	evaluative	evaluative
4. Even when the authors formally introduce \sigma and \omega in 4.2, it is still not clear that why both of them are used for modelling the success probability.	request	evaluative	evaluative
Hence, I would wholeheartedly recommend acceptance of this paper if the authors correct the factual errors (e.g. the claim of SO(3)-equivariance) and provide a clear discussion of the issues.	other	request	request
The concern raised by reviewer 3 is very important.	structuring	evaluative	evaluative
This paper proposes a Self-Modulation framework for the generator network in GANs, where middle layers are directly modulated as a function of the generator input z.	structuring	fact	structuring
However, it is not obvious that how to move from line 3 to line 4 at Eq 15.	evaluative	request	evaluative
- The authors perform numerous empirical experiments on several types of problems on various datasets (Digit, OFFICE,CELL) successfully showing how the MULANN can reduce the nasty effects of the adversarial domain discriminator and repulse (a fraction of) unlabeled examples from labeled ones in each domain.	evaluative	fact	evaluative
Efficiently learning a policy from visual inputs is an important research direction in RL.	evaluative	fact	fact
The authors did not comment on the computational overhead of training their LDA.	request	evaluative	evaluative
In any case, I very strongly suggest using a permutation approach to obtain the test threshold for HSIC, which is by far the most robust and reliable method.	none	request	request
- Table 1: Not sure why there is only one model that employs beam search (with beam size = 2) among all the comparisons. It looks strange.	evaluative	request	request

End of preview.

Dataset Content

Dataset Sources

The dataset is constructed by integrating two large-scale open resources: REVIEWER2 and SEA.

REVIEWER2 compiles papers and reviews from ICLR (2017-2023) via OpenReview and from NeurIPS through two sources: official proceedings (2016-2020) and OpenReview (2021-2022). SEA complements this collection by adding the most recent reviews from NeurIPS 2023 and ICLR 2024.

On the integrated corpus, we have performed fine-grained sentence-level category labeling, where each sentence is annotated with its specific functional type, to support more precise academic text mining and evaluation analysis tasks.

Downloads last month: 17

Total file size:

1.04 GB