Document dataset columns and remove example row

README.md

@@ -0,0 +1,126 @@
+---
+language:
+  - en
+task_categories:
+  - text-classification
+size_categories:
+  - 10K<n<100K
+tags:
+  - peer-review
+  - iclr
+  - rebuttal
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: train.csv
+      - split: test
+        path: test.csv
+---
+
+# ICLR Reviews
+
+A review-level dataset of ICLR 2024 and 2025 submissions, initial reviews, author rebuttals, and contextual metadata from OpenReview. The prediction target is whether a reviewer's final rating differs from their initial rating.
+
+Each row represents one official review with both initial and final ratings available and at least one retained author response in that review's discussion thread. Multiple rows can refer to the same paper.
+
+## Files and splits
+
+| File | Rows | Description |
+|---|---:|---|
+| `train.csv` | 44,126 | Training reviews. |
+| `test.csv` | 11,099 | Held-out reviews. |
+
+There are 55,225 rows in total. The source split is approximately 80/20 by paper, so reviews of the same paper stay together. The labels retain their natural, unbalanced distribution.
+
+## Columns
+
+Blank CSV fields represent missing or unavailable values. Binary indicators use `1` for true and `0` for false unless missing. Timing values are in fractional days, rounded to three decimal places. Within-paper ranks are one-based; ties share the smaller rank.
+
+### Identifiers and paper text
+
+| Column | Description |
+|---|---|
+| `paper_id` | OpenReview identifier of the paper. |
+| `review_id` | OpenReview identifier of the official review; identifies the review, not the reviewer. |
+| `year` | ICLR conference year: 2024 or 2025. |
+| `title` | Paper title. |
+| `abstract` | Paper abstract. |
+| `keywords` | Paper keywords, joined with semicolons. |
+
+### Review, responses, and target
+
+| Column | Description |
+|---|---|
+| `initial_review` | Review text after heuristic removal of paragraphs or sentences reporting post-rebuttal updates; long text is shortened as described below. |
+| `initial_rating` | Reviewer's initial numeric rating, using the source conference year's rating scale. |
+| `rebuttal` | Retained author messages in this review's thread, joined in chronological order with blank lines. |
+| `general_response` | Retained paper-level author messages, rather than responses specific to this review thread; may be blank. |
+| `n_rebuttal_messages` | Number of retained author messages in the review thread. |
+| `rebuttal_length` | Character count of the joined rebuttal text before shortening. |
+| `log_rebuttal_length` | Natural logarithm `ln(1 + rebuttal_length)`, rounded to four decimal places. |
+| `review_length` | Character count of the cleaned initial review before shortening. |
+| `label` | `changed` if the final rating differs from the initial rating; otherwise `unchanged`. Both increases and decreases count as changes. The final rating is not included as an input column. |
+
+### Timing and discussion context
+
+In the descriptions below, "first" and "last" rebuttal refer to the first and last **retained author messages** in the review thread. "Release" means the scheduled release of reviews to authors, not the paper's publication date. "Deadline" means the scheduled author-response deadline, including the extension for ICLR 2025.
+
+| Column | Description |
+|---|---|
+| `rebuttal_delay_days` | Days from creation of the official review to the first retained rebuttal. |
+| `review_days_before_release` | Days from creation of the review to the scheduled review release. |
+| `review_order_in_paper` | Chronological rank of this review's creation time among the paper's source reviews. |
+| `rebuttal_days_after_release` | Days from review release to the first retained rebuttal. |
+| `rebuttal_days_before_deadline` | Days from the first retained rebuttal to the author-response deadline. |
+| `last_rebuttal_days_before_deadline` | Days from the last retained rebuttal to the author-response deadline. |
+| `rebuttal_span_days` | Days between the first and last retained rebuttal; zero for a single message. |
+| `rebuttal_order_in_paper` | Rank of this thread's first retained rebuttal among first retained rebuttals for the paper. |
+| `rebuttal_in_extension` | Whether the first retained rebuttal was posted after the original ICLR 2025 deadline. Always zero for ICLR 2024, which has no extension threshold in this dataset. |
+| `general_response_before_rebuttal` | Whether any retained paper-level author response preceded this thread's first retained rebuttal. |
+| `n_other_reviewers_replied_before_rebuttal` | Number of other review threads for the same paper whose reviewer had replied before this thread's first retained rebuttal. |
+
+Positive "days before" values indicate an event before the reference time; negative values indicate an event after it.
+
+### Other reviews of the same paper
+
+These statistics are computed from the paper's source reviews, including reviews that may not become rows in the exported dataset. Statistics prefixed with `others_` exclude the current review and use available initial ratings.
+
+| Column | Description |
+|---|---|
+| `n_reviews` | Number of source reviews for this paper with an available initial rating, including the current review. |
+| `others_mean_initial_rating` | Mean initial rating of the other reviews, rounded to three decimal places; blank if none are available. |
+| `others_sd_initial_rating` | Population standard deviation of the other initial ratings, rounded to three decimal places; blank if fewer than two are available. |
+| `others_min_initial_rating` | Minimum initial rating among the other reviews; blank if none are available. |
+| `others_max_initial_rating` | Maximum initial rating among the other reviews; blank if none are available. |
+
+### Author and preprint metadata
+
+| Column | Description |
+|---|---|
+| `num_authors` | Number of authors recorded in the paper metadata. |
+| `author_affiliations` | Distinct recorded affiliation strings, joined with semicolons. |
+| `n_affiliations` | Number of distinct recorded affiliation strings; blank when affiliations are unavailable. |
+| `has_industry_author` | Whether an affiliation matches a predefined list of industry organization names. This is a heuristic, not a verified employment classification; blank when affiliations are unavailable. |
+| `arxiv_categories` | Recorded arXiv subject-category codes, generally separated by spaces; blank when unavailable. |
+| `arxiv_before_review` | Whether the recorded arXiv version-one timestamp precedes creation of this particular review. A zero can also reflect an unavailable arXiv timestamp; it does not prove that no earlier preprint existed. |
+
+## Text processing and limitations
+
+Author responses are retained only before the reviewer's first reply in the thread. If the reviewer never replies, that cutoff is unbounded. Within each author-message stream, retention also stops at the first message heuristically acknowledging an already changed rating. Paper-level responses use the same reviewer-reply cutoff. Initial-review text is heuristically cleaned to remove post-rebuttal updates.
+
+Long text retains its beginning and end, with an omission marker between them. The retained-text budgets are 8,000 characters for `initial_review`, 16,000 for `rebuttal`, and 8,000 for `general_response`; the omission marker adds a small number of characters beyond these budgets. Length features are measured before shortening.
+
+These filters reduce explicit outcome leakage but cannot guarantee its absence. In particular, response-span and last-response timing features partly reflect when, or whether, the reviewer replied. Metadata may be incomplete, affiliation matching is heuristic, and rating scales may differ across conference years. A rating change is an observed association, not evidence that the rebuttal caused the change or improved the paper. This dataset covers selected reviews with author responses, not every ICLR submission or review.
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("Harini/iclr_reviews", token=True)
+train = dataset["train"]
+test = dataset["test"]
+```
+
+For private access, authenticate with a Hugging Face account authorized to read this dataset. CSV readers should support quoted multiline text fields.

example_row.txt

@@ -1,84 +0,0 @@
-paper_id: 014CgNPAGy
-review_id: 3FOfBFYrdO
-year: 2024
-title: On the Role of Momentum in the Implicit Bias of Gradient Descent for Diagonal Linear Networks
-
-abstract:
-Momentum is a widely adopted and crucial modification to gradient descent when training modern deep neural networks. In this paper, we target on the regularization effect of momentum-based methods in regression settings and analyze a popular proxy model, diagonal linear networks, to precisely characterize the implicit bias of  heavy-ball (HB) and Nesterov's method of accelerated gradients (NAG). We show that, HB and NAG exhibit different implicit bias compared to GD for diagonal linear networks, which is different from the one for  classic linear regression problem where momentum-based methods share the same implicit bias with GD. Specifically, the role of momentum in the implicit bias of GD is twofold. On one hand, HB and NAG induce extra initialization mitigation effects similar to SGD that are beneficial for generalization of sparse regression. On the other hand, besides the initialization of parameters, the implicit regularization effects of HB and NAG also depend on the initialization of gradients explicitly, which may not be benign for generalization. As a consequence, whether HB and NAG have better generalization properties than GD jointly depends on the aforementioned twofold effects determined by various parameters such as learning rate, momentum factor, data matrix, and integral of gradients. Particularly, the difference between the implicit bias of GD and that of HB and NAG disappears for small learning rate. Our findings highlight the potential beneficial role of momentum and can help understand its advantages in practice from the perspective of generalization.
-
-keywords: GD; momentum; implicit bias; linear networks
-
-initial_review:
-summary:
-The paper studies the implicit regularization effect of the momentum methods through their continuous-time approximations in the case of diagonal networks. The paper discusses the role of initialization and the gradient of the loss at initialization on the recovered implicit bias. The theory is supported by experiments.
-
-strengths:
-The paper addresses a well-motivated problem. Given the widespread use of momentum in training deep networks, gaining insights into its implicit regularization and convergence, even for simple non-convex models, is essential.
-
-weaknesses:
-I believe that Theorem 1 is incorrect or at least it is incorrectly stated. 
-
-The paper studies momentum ODE but the problem is in the analysis, i.e., proof of Proposition 3 and Theorem 1, the $\eta^2$ terms are ignored even after the modelling with an ODE. But the result stated does not state that the implicit bias holds only under this approximation. Hence, the statement in its current form is incorrect.  I am happy to engage in further discussion if the authors bring other points of discussion. 
-
-I also do not agree with the proof technique. a) either start with a discrete time algorithm and ignore the higher order terms to recover the potential which it implicitly minimizes or b) use the continuous time counterpart and state the implicit bias for this without any further approximation. In the approach followed by the paper it is not clear what is the order of approximation of the result (the theorem does not even state that it is an approximate result). There are two errors: one stemming from discretization and the other from neglecting second-order terms. It is not evident how close this trajectory to the one described in Theorem 1. However, I strongly feel that the Theorem 1 should be rewritten and it is not acceptable in its current form. I
-
-questions:
-already discussed in the weakness part.
-
-initial_rating: 3
-
-rebuttal:
-Our response
-We thank the reviewer for the valuable comments and questions. Below we address your concerns.
-
-### **Weaknesses Part**
-----
-
-**(1). On neglecting second-order terms.**
-
-We would like to clarify that neglecting the second-order terms for the ODE is eligible. This is because the momentum ODE is established to the order of $\eta$, i.e., the momentum ODE in Eq.(4) is in fact $\alpha \ddot{\beta} + \dot{\beta} + \nabla L / (1 - \mu) = \mathcal{O}(\eta^2)$. Thus all terms of order higher than $\eta$ (including the order $\eta^2$) appearing in anywhere else should also be dropped. To be consistent with this convention, the $\eta^2$ terms are ignored in our proof. For example, from Eq.(4) we have
-$ \eta \dot{\beta} = - \eta\frac{\nabla L}{1 - \mu} + \mathcal{O}(\eta^2),$  which is eligible (it states that momentum adds a perturbation of the order $\eta$ to re-scaled gradient flow) and can also be derived from the discrete update rules by following exactly the same approach of the proof for Proposition 1 and ignoring $\eta^2$ terms in the discrete relation Eq.(14).
-
-We mentioned this point in the proof and we thank the reviewer for pointing out that we missed this in the statement of theorems. We fix this in the revision by stating clearly in Proposition 1 that we ignore all terms of the order higher than $\eta$.
-
-One can try to generalize the analysis to include terms of the order higher than $\eta$ by using the momentum ODE with higher order terms, e.g., Eq.(13) on page 11 of Kovachki and Stuart (2021) that includes arbitrarily higher order terms, and following our current approach where one should now preserve these higher order terms. Since most momentum-based modellings of momentum-based methods are truncated to the order of $\eta$, we also focus on this choice. 
-
-----
-**(2). On the order of the continuous approximation of HB and NAG and inaccurate statement of Theorem 1.**
-
-The continuous counterpart of HB and NAG used in our paper (Eq.(4)) is $\mathcal{O}(\eta)$ approximate continuous version of discrete HB and NAG, as shown in Kovachki and Stuart (2021). We thank the reviewer for the suggestion of inaccurate statement of Theorem 1, and we fix this in the revision by carefully and clearly stating the current implicit bias is for the $\mathcal{O}(\eta)$ approximate continuous version of HB and NAG, or more concretely HB and NAG flow, rather than directly for the discrete HB and NAG.
-
-We adopt the $\mathcal{O}(\eta)$ approximate continuous version of momentum-based methods since we aim to analyze the role of momentum in the implicit bias of widely-studied gradient flow, which is also an order $\mathcal{O}(\eta)$ approximate continuous version of GD, for diagonal linear networks, and the analysis should be conducted in the same order of approximation. The analysis can be naturally generalized to the order $\mathcal{O}(\eta^2)$ approximate continuous version of HB by replacing the $\mathcal{O}(\eta)$ approximate continuous version of HB (Proposition 1) with the $\mathcal{O}(\eta^2)$ ones in Kovachki and Stuart (2021) (Eq.(20)) and following similar approach as in the current work.
-
-----
-### **Reference**
-Kovachki and Stuart. Continuous Time Analysis of Momentum Methods. arXiv: 1906.04285, 2019.
-
-general_response: 
-n_rebuttal_messages: 1
-rebuttal_length: 3261
-log_rebuttal_length: 8.0901
-review_length: 1901
-rebuttal_delay_days: 14.893
-n_reviews: 4
-others_mean_initial_rating: 4.0
-review_days_before_release: 8.182
-review_order_in_paper: 4
-rebuttal_days_after_release: 6.711
-rebuttal_days_before_deadline: 6.08
-last_rebuttal_days_before_deadline: 6.08
-rebuttal_span_days: 0.0
-rebuttal_order_in_paper: 3
-rebuttal_in_extension: 0
-general_response_before_rebuttal: 0
-n_other_reviewers_replied_before_rebuttal: 0
-others_sd_initial_rating: 2.16
-others_min_initial_rating: 1
-others_max_initial_rating: 6
-num_authors: 2
-author_affiliations: University of Southampton
-n_affiliations: 1
-has_industry_author: 0
-arxiv_categories: 
-arxiv_before_review: 0
-label: unchanged