Pith. sign in

REVIEW 4 major objections 5 minor 39 references

REDDIX-NET: A Novel Dataset and Benchmark for Moderating Online Explicit Services

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ReddiX-NET, a new benchmark of 8,146 labeled Reddit posts, lets large language models sort online sexual services into six behavioral categories, with the best tested model averaging about 92 percent accuracy.

desk verdict A potentially useful dataset idea that is currently impossible to evaluate because the data is withheld and the evaluation protocol contradicts itself. read the letter →

arxiv 2505.23231 v1 pith:U6K3WZJH submitted 2025-05-29 cs.CY

MSVPJ Sathvik , Manan Roy Choudhury , Rishita Agarwal , Sathwik Narkedimilli , Vivek Gupta This is my paper
classification cs.CY
keywords onlinesexualservicescontentmoderationbenchmarkdatasetRedditanalysisLLMclassificationsentimenttemporalNSFW
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces REDDIX-NET, a benchmark dataset of 8,146 Reddit posts labeled into six classes of online sexual services, and argues this fills a gap left by traditional NSFW filters. It claims first-mover status for such a dataset, and reports that large language models like Claude 3.5 Haiku, GPT-4, and Gemini 1.5 Flash can classify these posts with average accuracies of 92.49 percent, 83.72 percent, and 83.85 percent respectively. If the claim holds, moderation systems gain a reusable test set that captures transactional cues, coded language, and service-specific behavior rather than explicit imagery alone. The dataset also supports sentiment, comment, and temporal analyses that the paper offers as evidence of distinct engagement patterns worth moderating.

What carries the argument

The load-bearing object is REDDIX-NET itself: a hand-labeled corpus whose labels were set by three annotators (two authors and one other) using six categories that the paper defines in Section 3.2. The paper reports Krippendorff's alpha of 0.696, Fleiss kappa of 0.608, and pairwise Cohen's kappa values of 0.554, 0.681, and 0.74 across annotator pairs, treating these as evidence that the labels are reliable. The evaluation machinery is a zero-shot prompt asking each LLM to return exactly one of the six category names (or 'Uncategorizable'), with the model's output compared against the annotation-derived labels using accuracy, precision, F1, MSE, MAE, and Jensen-Shannon divergence; a clustering fallback is used for posts the models refuse to categorize. Secondary machinery includes BERT-based and LLM-based sentiment/emotion/tone extraction, a 19-bucket comment classification scheme, and timestamp analysis of posting activity.

What would settle it

Take a random sample of about 500 posts from REDDIX-NET and have a fresh annotation team, blind to the original labels, assign them to the same six categories; if pairwise agreement with the published labels falls below a conventional kappa threshold like 0.4, the benchmark's ground truth is not reproducible. A complementary check is to count posts whose content plausibly fits two or more categories; if that fraction is substantial, the single-label accuracy scores in Table 2 do not measure what they claim.

Watch

Extended reading notes

Core claim

The paper claims to be the first to construct a dataset specifically for detecting online sexual services, rather than merely flagging NSFW content. Its central evidence is REDDIX-NET, 8,146 posts collected from three large subreddits and manually labeled into six service classes: Content Creation and Sales, Couples and Group Interactions, Exhibitionism, Miscellaneous Fun, Physical Services, and Virtual Services. Using a single-label prompting harness that feeds each post's title, image description, and comments to state-of-the-art large language models, the paper reports average classification accuracies of 92.49 percent for Claude 3.5 Haiku, 83.85 percent for Gemini 1.5 Flash, 83.72 percent for GPT-4, and comparable figures for other models, alongside large variations in per-category F1. It further claims that sentiment, emotion, comment, and temporal analyses of the same dataset reveal distinct engagement patterns, such as peak activity between 12 and 19 hours and category-specific emotional tones, that could inform moderation and intervention timing.

Load-bearing premise

The benchmark's ground truth is the set of labels assigned by three annotators to six categories that are assumed exhaustive and mutually exclusive, an assumption the paper itself contradicts by calling the task multi-label in its error-analysis appendix.

Editorial extensions

If this is right

  • Moderation systems gain a shared, reusable test set for detecting solicitation of sexual services, one that includes transactional cues and coded language rather than only explicit imagery.
  • The per-category F1 scores provide a direct ranking of which service classes are hardest for current LLMs, showing Exhibitionism and Couples and Group Interactions as the main failure points across all tested models.
  • The temporal findings imply that moderation resources could be concentrated in the 12–19 hour window when posting and commenting activity peaks on these subreddits.
  • The reported inter-annotator agreement values give future users a quantitative sense of label noise, letting them calibrate how much classification error comes from the models versus the dataset itself.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial: Because the paper's prompt demands a single label while its error analysis treats the task as multi-label, we infer the dataset would support higher per-category quality if released with multi-label annotations, especially for classes like Virtual Services and Content Creation that the error analysis shows confusingly overlap.
  • Editorial: The excellent F1 reported for Miscellaneous Fun (100 percent for Claude, 83.64 percent for Gemini) is likely an artifact of that class having only 105 posts, far fewer than the others; we would expect performance to drop if the class were balanced.
  • Editorial: The dataset's value for real-world moderation would be tested by applying the same prompts to posts from other platforms such as Twitter/X, TikTok, or Snapchat; the paper itself notes the dataset is Reddit-specific, so we expect accuracy to fall without additional annotation.
  • Editorial: The paper's first-mover claim could be checked with a systematic search of trafficking- and sex-work-related datasets; if earlier labeled sexual-service solicitation datasets exist, the novelty narrows to the six-category taxonomy and the social-media posting context.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces REDDIX-NET, a dataset of 8,146 Reddit posts collected from three subreddits and annotated into six service categories (CCS, CGI, Ex, MF, PS, VS). The authors benchmark six large language models on post classification, reporting precision, F1, MSE, MAE, JSD, and accuracy, and then present sentiment, emotion, tone, comment, and temporal analyses. The headline claim is that models such as GPT-4 and Gemini 1.5 Flash achieve promising classification results, and the authors assert that this is the first dataset on online sexual services.

Significance. If the dataset were publicly available and the evaluation protocol were sound, REDDIX-NET could fill a real gap in content-moderation research by moving beyond binary NSFW filters toward fine-grained categories of online sexual services. The paper contains substantial collection and annotation effort, statistical significance tests, and a wide range of exploratory analyses. However, the internal inconsistencies in the task definition, the arithmetically implausible metric values, and the absence of accessible data or code mean that the central empirical claims cannot currently be verified or reused. The significance is therefore not established as submitted.

major comments (4)
  1. [Appendix-4, Appendix-1, Table 2] The task definition is internally contradictory. Appendix-4 explicitly states that the task is multi-label and gives a gold example with labels (Ex, VS, PS), while Prompt 1 in Appendix-1 instructs the model to return exactly one of six categories or 'Uncategorizable.' Table 2 then reports per-category precision, F1, and accuracy without explaining how a single-label prediction is scored against multi-label ground truth. Under these conditions, the reported metrics are not well-defined and the benchmark results cannot be independently reproduced. Please specify whether the task is single-label or multi-label, and if multi-label, either obtain multi-label predictions or state the exact matching rule used for scoring.
  2. [Table 2] Several entries in Table 2 are arithmetically impossible under standard definitions. For Claude 3.5-Haiku in the MF row, Precision=100% and F1=100% imply no false positives and no false negatives, which under the per-category TP/FP/FN/TN definitions in Appendix-4 would require Accuracy=(TP+TN)/(TP+FP+FN+TN)=100%, yet the table reports Acc=78.48%. Similar inconsistencies appear in other rows, for example GPT-4 MF with Pre=57.14 and F1=72.02 cannot yield Acc=84.60 under one-vs-rest counting. The 'Acc' column therefore appears to measure something other than standard classification accuracy, and the paper must provide exact formulas and recompute the table.
  3. [Appendix-5, Section 3.2] The gold-label quality is not sufficiently established. Table 6 reports Cohen's kappa of 0.554 for the (1,2) annotator pair and Fleiss kappa of 0.608, which are not 'high' agreement as claimed in Section 3.2. With all annotators being authors or members of the same organization, and with no independent adjudication, released annotation guidelines, or external validation, the ground truth that every LLM score depends on is fragile. Please report per-category agreement, describe adjudication procedures, and provide an external annotation check or at least a detailed label-error analysis.
  4. [Ethics Statement, Abstract] The dataset is not available to reviewers or readers. The Ethics Statement says that access will be granted only after submitting a research-use form and that only a curated subset will be released, but no link, repository, or concrete access mechanism is provided anywhere in the manuscript. A benchmark paper cannot be evaluated or reused without at least a review copy of the data and annotations, or a clearly documented and currently available access procedure. Please include these materials or explain why they cannot be provided.
minor comments (5)
  1. [Section 3.3] The text says 'The word count across all models reaches 557,764,' but the surrounding context and Table 1 refer to categories, not models; this should be corrected.
  2. [Appendix-10] The reported average accuracy for Claude is 92.49%, whereas Table 2 averages to approximately 82.49%; this inconsistency should be corrected before the comparison with prior work is used to support claims of superior performance.
  3. [Section 4.1] The handling of 'Uncategorizable' predictions by clustering is described only at a high level; please specify how many posts were uncategorized, how clusters were formed, and how cluster labels were assigned, since this affects the reported metrics.
  4. [Table 2 caption, Appendix-1] The table caption and Appendix-1 list metric names but give no formulas; please provide precise definitions for MSE, MAE, JSD, and 'Accuracy,' and state clearly whether they are computed per-category, per-post, or as distribution-level measures.
  5. [Section 2, Contribution (a)] The claim of being 'the first to develop a dataset on online sexual services' should be qualified in light of the closely related escort-ad and trafficking datasets cited in Section 2; as written, the novelty claim is too broad.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: REDDIX-NET is a self-contained annotated benchmark; the LLM evaluation is zero-shot against human labels, and the multi-label/single-label inconsistency is a correctness risk rather than a circular derivation.

full rationale

The paper's central contribution is a labeled dataset: ground-truth labels come from human annotation in Section 3.2, not from the models being benchmarked. Section 4.1 prompts LLMs to classify posts into the predefined categories, and Table 2 reports agreement between those zero-shot outputs and the human labels; no parameter is fitted to the evaluation labels, and no prediction is re-derived from the metric being reported. The paper does not invoke a load-bearing self-citation: the prior works cited (Ibanez and Suthers, Diaz and Panangadan, Wang et al.) are external and are used for comparison or motivation, not to justify the dataset's validity. The Appendix-4 statement that the task is multi-label, with a worked example showing multiple ground-truth labels, conflicts with the single-label prompt in Appendix-1 and with Table 2's per-category metrics; this is an internal evaluation-protocol inconsistency that makes the benchmark numbers ambiguous, but it does not make the results equivalent to the inputs by construction. Similarly, Appendix-1's boilerplate mention of fine-tuning hyperparameters conflicts with the main text's zero-shot description, but the paper never specifies that the evaluation labels were used for training, so the reduction required for a fitted-input/called-prediction finding is not exhibited. Because the dataset is defined by human annotation and the LLM scores are external comparisons against those annotations, the derivation chain is self-contained and no circular step is established.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper's claims rest on the validity of the annotator-defined taxonomy and the assumption that the scraped posts are real service offers. No free parameters are fitted; the LLM evaluations are zero-shot, and the statistical tests are standard. The key unvalidated input is the ground truth itself, created by the authors, with moderate inter-annotator agreement.

assumptions (3)
  • domain assumption The six hand-defined categories (CCS, CGI, Ex, MF, PS, VS) are exhaustive and mutually exclusive for online sexual-service posts.
    Section 3.2 defines the six categories and treats them as the label set, but the evaluation prompt includes an 'Uncategorizable' option and Appendix-4 says posts may belong to multiple categories, so the assumption is internally violated.
  • domain assumption The scraped Reddit posts are genuine offers of paid sexual services, not roleplay, spam, satire, or undercover content.
    Section 3.1 states the authors 'gathered posts offering services' without verifying transactional intent beyond the annotators' reading of the text.
  • domain assumption Zero-shot LLM classification on the prompt template is a valid proxy for real-world moderation performance.
    Section 4.1 uses the resulting F1 and accuracy scores to claim 'promising results', but no human-in-the-loop moderation baseline or deployment context is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REDDIX-NET: A Novel Dataset and Benchmark for Moderating Online Explicit Services." pith.science (2026). https://pith.science/paper/U6K3WZJH

@misc{pith2026250523231,
  author       = {Pith},
  title        = {Pith review of: REDDIX-NET: A Novel Dataset and Benchmark for Moderating Online Explicit Services},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U6K3WZJH}},
  note         = {Machine review of arXiv:2505.23231}
}
read the original abstract

The rise of online platforms has enabled covert illicit activities, including online prostitution, to pose challenges for detection and regulation. In this study, we introduce REDDIX-NET, a novel benchmark dataset specifically designed for moderating online sexual services and going beyond traditional NSFW filters. The dataset is derived from thousands of web-scraped NSFW posts on Reddit and categorizes users into six behavioral classes reflecting different service offerings and user intentions. We evaluate the classification performance of state-of-the-art large language models (GPT-4, LlaMA 3.3-70B-Instruct, Gemini 1.5 Flash, Mistral 8x7B, Qwen 2.5 Turbo, Claude 3.5 Haiku) using advanced quantitative metrics, finding promising results with models like GPT-4 and Gemini 1.5 Flash. Beyond classification, we conduct sentiment and comment analysis, leveraging LLM and PLM-based approaches and metadata extraction to uncover behavioral and temporal patterns. These analyses reveal peak engagement times and distinct user interaction styles across categories. Our findings provide critical insights into AI-driven moderation and enforcement, offering a scalable framework for platforms to combat online prostitution and associated harms.

Figures

Figures reproduced from arXiv: 2505.23231 by the authors.

Figure 1
Figure 1. Structure of the proposed dataset, categorizing online sexual services across six distinct categories and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of sentiment classifications across [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Category-wise impact proportion highlighting [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Visualization of temporal posts activity pat [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Correlation heatmaps depicting the interplay between sentiment, emotion, and tone across multiple dataset [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Distribution of tone and emotions across sentiment types for different classes in the dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The Confidence Score Distribution by Sentiment: Box Plot Comparison of Positive, Neutral, Mixed, and [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Sentiment Analysis using BERT Model In reviewing [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Total No. of Performers per Category and their Overlap [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Confusion matrix for Llama 3.3 70B [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 13
Figure 13. Figure 13: Confusion matrix for Gemini 1.5 Flash [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 15
Figure 15. Figure 15: Confusion matrix for Mistral 8x7B Class-by-Class Failure Analysis We have two types of data: the ground truth data and the LLM-predicted data. The ground truth data tells us which posts belong to which particular categories out of the six categories. We then compared …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

  1. [1]

    Physical Services: Posts offering in-person sexual services or physical prostitution

    work page

  2. [2]

    {image_description}

    Image Description: "{image_description}"

    work page

  3. [3]

    {comments}

    Comments: "{comments}" Your task is to classify this post into one of the following six categories of services related to online prostitution:↪→

    work page

  4. [4]

    Virtual Services: Posts offering virtual interactions such as video calls, virtual sex, or promoting platforms like OnlyFans

    work page

  5. [5]

    Exhibitionism: Posts showcasing exhibitionistic behavior, such as public displays or other forms of showcasing oneself

    work page

  6. [6]

    Content Creation and Sales: Posts promoting or selling photos, videos, or other content without direct interaction

    work page

  7. [7]

    Couples and Group Interactions: Posts seeking interactions involving couples, threesomes, or group scenarios

    work page

  8. [8]

    Uncategorizable

    Miscellaneous Fun/Exploration: Posts describing non-specific fun, exploration, or interactions that do not fall into the other categories.↪→ Carefully analyze the title, image description, and comments. Then, determine which of the six categories the post best fits into. Respond with only the category name. If the information is insufficient, respond with...

    work page

Show all 39 references
  1. [9]

    Sentiment: Positive, Neutral, or Negative, with a confidence score (0-1)

  2. [10]

    Emotion Classification: Identify the dominant emotion (e.g., joy, anger, sadness, surprise, etc.).↪→

  3. [11]

    Keywords: Extract the main keywords or phrases relevant to the context

  4. [12]

    Tone: Determine the tone (e.g., formal, casual, playful, persuasive, etc.). Output Format: • Sentiment: [label], Confidence: [score] • Emotion: [emotion] • Keywords: [keywords] • Tone: [tone] The provided sentiment analysis prompt instructs an expert AI to perform a comprehens...

  5. [13]

    Wow! You look absolutely stunning!

    Users Who Are Enjoying the Post and Its Contents (Engagement & Positive Sentiment) - "Wow! You look absolutely stunning!" - "Absolutely mesmerizing, I'm hooked!" - "Stunning visuals, keep up the great work!"

  6. [14]

    How much do you charge for this service?

    Users Who Are Demanding Such Services (Intent to Purchase/Engage) - "How much do you charge for this service?" - "Where can I reach you for more details?" - "Are you available for a private session?"

  7. [15]

    Can you do a video in a red dress?

    Users Who Are Requesting Specific Content (Content Demand Trends) - "Can you do a video in a red dress?" - "I’d love to see more dance moves from you!" - "Could you post more outdoor shoots?"

  8. [16]

    Is this actually you or just edited?

    Users Who Are Skeptical or Questioning Authenticity (Trust & Credibility Issues) - "Is this actually you or just edited?" - "Has anyone actually met her? Looks fake." - "Seems too polished—are these authentic?"

  9. [17]

    She’s super professional and amazing to work with!

    Users Who Are Providing Reviews & Recommendations (Word-of-Mouth & Service Feedback) - "She’s super professional and amazing to work with!" - "Had a great time, she’s very professional!" - "Overpriced, not worth it."

  10. [18]

    Isn’t this kind of thing banned here?

    Users Who Are Discussing Legality & Ethics - "Isn’t this kind of thing banned here?" - "Should this even be allowed on this platform?" - "I’m concerned about the legality of this content."

  11. [19]

    Check out my profile if you like this!

    Users Who Are Competing or Self-Promoting Services - "Check out my profile if you like this!" - "I offer exclusive content at a discount!" - "I can do this for half the price. DM me!"

  12. [20]

    Can you do this for \$40 instead?

    Users Who Are Negotiating Prices or Services - "Can you do this for \$40 instead?" - "Is there any discount if I book multiple sessions?" - "How about a special rate for returning customers?"

  13. [21]

    I paid but never got my order!

    Users Who Are Complaining About Payment or Delivery Issues - "I paid but never got my order!" - "She stopped replying after I sent the payment!" - "This is a scam, don’t fall for it!"

  14. [22]

    Send me something for free first!

    Users Who Are Making Unsolicited Requests or Harassment - "Send me something for free first!" - "I’ll find you if you don’t reply!" - "Do this for me, or else!"

  15. [23]

    Teri nazaron mein vo jadu hai

    Comments that are Multi-Lingual - "Teri nazaron mein vo jadu hai." - "Include comments from languages which are not English." - "acha hai"

  16. [24]

    Show me an online act that blends erotic fantasy with a violent edge

    Comments that are Fantasy and Violent Demands - "Show me an online act that blends erotic fantasy with a violent edge." - "I demand you to enact a dark fantasy scene with intense aggression."

  17. [25]

    Your posts always brighten my day!

    Comments that are Emotional Support - "Your posts always brighten my day!" - "I appreciate your openness; it helps me feel less alone."

  18. [26]

    Is this really you or just an impersonator?

    Comments that are Verification and Identity Inquiries - "Is this really you or just an impersonator?" - "Can you prove that this is your real account?"

  19. [27]

    Check this link out for exclusive content: [external link]

    Comments that are External Link Sharing - "Check this link out for exclusive content: [external link]" - "Visit my page for more: [link]"

  20. [28]

    Can you send me private pics? I’ll pay extra

    Comments that are Sexual Propositions / Explicit Requests - "Can you send me private pics? I’ll pay extra." - "Do you do custom videos with nudity?"

  21. [29]

    I’m the best at what I do, no one compares!

    Comments that are Self-Assertive/Confidence Expressions - "I’m the best at what I do, no one compares!" - "I always get what I want, and this is no different."

  22. [30]

    I’m not sure what to think about this

    Comments that are Ambiguous or Multi-Response Comments - "I’m not sure what to think about this..." - "Interesting... I wonder what’s really going on." For each comment provided, classify it into exactly one of the above categories and return the output as a JSON object with e...

  23. [31]

    online sessions

    VS (Virtual Services) • High FN for some models: Subtle language such as “online sessions” or “DM for details” sometimes goes undetected, increasing FN. • FP: Certain LLMs mislabel purely conversational or flirtatious content as “VS” even when no explicit service is being offered

  24. [32]

    meet and greet,

    PS (Physical Services) • Lower FNs overall, suggesting that explicit phrases referencing physical meetups or location-based transactions are easier for the LLMs to pick up. • Moderate FP: Posts hinting at offline gatherings for nonsexual contexts (e.g., “meet and greet,” “hang...

  25. [33]

    miscellaneous fun

    MF (Miscellaneous Fun) • Frequent Confusions: Many models confuse MF with either VS or PS because “miscellaneous fun” can overlap with playful or euphemistic language. This results in both FP (flagging other categories as MF) and FN (failing to catch truly offbeat or playful s...

  26. [34]

    public,” “show,

    Ex (Exhibitionism) • High FN: Models often fail to classify content as Ex if it doesn’t contain explicit words like “public,” “show,” or “watch,” thus missing subtle references (e.g., partial mention of webcam exhibition). • FP: Overzealous classification of normal adult posts...

  27. [35]

    We’re looking for new friends

    CGI (Couples and Group Interactions) • Contextual Mistakes: Many false negatives occur when more than two people are mentioned, but the nature of the post isn’t strictly a couples/group sexual context (e.g., “We’re looking for new friends” can be mistaken for group sexual acti...

  28. [36]

    collab,” “premium account,

    CCS (Content Creation and Services) • High Overlap with VS: Offers for camming, custom videos, or phone services sometimes get mislabeled as plain Virtual Services. This confusion causes FP under CCS and FN under VS—or vice versa—depending on the model’s bias. • Industry Slang...

  29. [37]

    Data Collection The data for this study was collected from three specific subreddits identified as primary hubs for discussions related to sexual services. Data extraction was performed using the Reddit API, facilitated by the PRAW (Python Reddit API Wrapper) library, which en...

  30. [38]

    Hi," "Hello!

    Data Cleaning The initial dataset underwent a cleaning process to eliminate irrelevant or extraneous content. Posts and comments deemed non-substantive, such as greetings (e.g., "Hi," "Hello!"), were removed to ensure the dataset focused solely on meaningful exchanges related ...

  31. [39]

    Posts containing visible faces were excluded from the dataset, as most posts naturally blurred such identifying features

    Data Preprocessing To protect the anonymity of individuals involved, several preprocessing steps were implemented. Posts containing visible faces were excluded from the dataset, as most posts naturally blurred such identifying features. Additionally, all usernames and Reddit I...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.