REVIEW 3 major objections 6 minor 1 cited by
ToxSyn-PT: A Synthetic Fine-Grained Dataset of Minority-Targeted Toxic Language in Portuguese
T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that ToxSyn, a new 53,274-sentence synthetic Portuguese corpus, is the first resource of its kind to enable fine-grained multi-label hate-speech detection across nine protected minority groups, and that cross-domain…
desk verdict A genuinely useful new Portuguese hate-speech dataset with credible human validation, but the headline 'mutual catastrophic failure' claim is only demonstrated on synthetic/translated test sets for the social-media-to-minority direction. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the controllable four-stage generation pipeline: (1) a 160-sentence seed set of toxic and non-toxic examples about Black people and women; (2) few-shot prompting of GPT-4o Mini with 26 to 28 generation templates to expand coverage to nine protected groups, yielding 10,790 sentences; (3) dual paraphrase augmentation that rewrites each sentence through toxic and non-toxic transformations, tripling the corpus and creating hard counterexamples; and (4) an enrichment round that re-prompts from the augmented corpus and adds implicit-hate templates such as Ambiguous Prejudice and Justification Prejudice, yielding 50,074 minority-targeted sentences. A separate module adds 3,200 neutral samples. The pipeline is what carries the argument: because it generates balanced labels and explicit non-toxic counterexamples, it makes group-specific training and evaluation possible in Portuguese for the first time.
What would settle it
Train BERTimbau on size-matched, class-balanced subsets (equal toxic and non-toxic counts, equal total examples) of ToLD-BR, HateBR, OLID-BR, and ToxSyn, with several random seeds, and rerun the cross-domain evaluation; if out-of-domain toxic recall stops collapsing below 0.60, the paper's mutual-failure claim is an artifact of unequal training conditions rather than evidence that social-media and minority-targeted hate are distinct tasks.
Extended reading notes
Core claim
ToxSyn's central claim is that it is the first Portuguese corpus enabling fine-grained, multi-label detection of hate directed at nine protected groups-Black people, women, LGBTQIA+ individuals, Indigenous Brazilians, Muslims, Jewish people, elderly people, people with disabilities, and immigrants-because it is the first to include non-toxic counterexamples (benign or supportive sentences about those groups) alongside toxic ones. The corpus totals 53,274 sentences: 50,074 minority-targeted samples and 3,200 neutral samples, each annotated for binary toxicity, target group, and discourse type. The paper's headline experimental result is a mutual, cross-domain generalization failure: a BERTimbau model fine-tuned on ToLD-BR finds only 10 of 2,472 toxic ToxSyn test sentences, the best social-media-trained model (OLID-BR) recalls only about 20% of minority-focused toxicity, and a ToxSyn-trained model's toxic recall drops to 0.17 on ToLD-BR and 0.51 on HateBR. Because Macro F1 stays between 0.42 and 0.69 across these collapses, the paper argues that summary metrics can completely mask the failure to detect hate.
Load-bearing premise
The paper's central experiment assumes that differences in training-set size, class balance, and label definitions across the four datasets do not drive the measured recall collapse, so the failure can be blamed on domain shift rather than on the experimental setup.
Editorial extensions
If this is right
- A ToxSyn-trained model reaches 0.69 Macro F1 and 0.77 toxic-class recall on translated ToxiGen, evidence that minority-focused synthetic data transfers to another minority-focused dataset.
- Social-media-trained Portuguese classifiers cannot be assumed to detect minority-targeted hate: the best such model recalls only one in five minority-targeted toxic sentences.
- Macro F1 should not be the headline metric for hate-speech evaluation, since it masked near-zero toxic recall in these experiments.
- The four-stage pipeline provides a template for building similarly balanced, fine-grained hate-speech corpora in other low-resource languages.
Reading between the lines
- Testable extension: mixing ToxSyn with social-media data into a single hybrid training set could produce one classifier that covers both registers, though the paper proposes this idea as future work without testing it.
- An untapped implication is that the discourse-type labels (victim blaming, ambiguous prejudice, positive negation, etc.) could be used to train models to classify how hate is expressed, not just whether it is hate; the paper mentions this as an open direction.
- If the mutual failure is truly domain-driven, the paper implies that general-purpose Portuguese hate-speech systems will continue to under-protect minority groups unless they are trained on minority-focused data or explicitly adapted across domains.
- The paper discards roughly 8% of generated sentences during target normalization, suggesting some generated text expresses multiple or ambiguous targets; modeling intersectional or multi-target hate could recover signal that ToxSyn's single-target schema sets aside.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces ToxSyn (arXiv title: ToxSyn-PT), a synthetic Portuguese corpus of 53,274 sentences designed for multi-label hate-speech detection across nine protected minority groups, with balanced toxic/non-toxic labels, discourse-type annotations, and non-toxic counterexamples. The corpus is built by a four-stage pipeline: seed examples, GPT-4o-mini prompt-based expansion, paraphrase augmentation, and enrichment, followed by a neutral-sample module. A 5,208-sample test set was annotated from scratch by four native Portuguese speakers, with roughly 7% of the generated toxicity labels changed and fewer than 1% of target labels changed. The authors fine-tune BERTimbau on ToxSyn and on existing Portuguese datasets (ToLD-BR, HateBR, OLID-BR) and report cross-domain macro-F1 and toxic-class recall. They find that social-media-trained models have very low toxic recall on ToxSyn and translated ToxiGen, and that the ToxSyn-trained model has low toxic recall on the social-media datasets, concluding that there is a catastrophic mutual generalization failure and that macro-F1 masks this failure.
Significance. If the central empirical claims hold, ToxSyn is a valuable new resource: it is the first large-scale Portuguese corpus with fine-grained multi-label minority-target annotations plus the non-toxic counterexamples needed to separate hate from benign discussion. The pipeline is described concretely, the dataset is publicly released, and the human-validation protocol is a genuine strength because the gold-standard test set labels were produced independently of the model. The cross-domain finding, if robust, would be an important caution about summary metrics in hate-speech evaluation. The main risk is that the headline claim of 'catastrophic, mutual generalization failure' currently rests on comparisons that are not sufficiently controlled and on synthetic/translated test sets for one direction.
major comments (3)
- [§4.3, Table 4] The cross-domain comparison is not controlled for training-set size, class balance, or label-definition differences, all of which are load-bearing for the claim of a 'catastrophic mutual generalization failure.' The models are fine-tuned on full training sets with a single fixed seed, and the datasets differ by orders of magnitude in minority-targeted sample count (e.g., Table 2: ToLD-BR 124, HateBR 727, OLID-BR 1,390, ToxSyn 50,074). A model trained on a much smaller or more imbalanced corpus could show low out-of-domain recall for reasons unrelated to minority-targeting per se. The authors should add matched-size training runs (e.g., subsampling ToxSyn to the size of each social-media corpus, or augmenting social-media training sets to ToxSyn's size), repeat with multiple seeds, and report variance before attributing the observed recall collapse to domain shift.
- [§4.3, with §4.2 and §8] The social-media-trained models' failure on 'minority-focused data' is demonstrated only on ToxSyn itself and on translated ToxiGen, both of which are machine-generated or machine-translated. As the paper notes, ToxSyn has 'repetitive phrasing and distributional artifacts' (Section 8) and translated ToxiGen has 'unnaturalized phrasing and culturally mismatched slurs' (Section 4.2). Low recall on these sets could therefore reflect out-of-distribution synthetic text rather than a genuine property of minority-targeted hate. The one real minority-targeted benchmark, Portuguese Merged (Section 4.2), is never used to report binary toxic recall for the social-media-trained models. The authors should report toxic recall and F1 for all models on Portuguese Merged, and/or on the human-validated subset of ToxSyn, to determine whether the recall collapse persists on human-authored minority-targeted text.
- [§5, Discussion] The 'mutual' claim is asymmetric in the current evidence. The ToxSyn-trained model is tested on real social-media corpora (ToLD-BR, HateBR, OLID-BR), but the social-media-trained models are tested on minority-focused sets that are both synthetic/translated. Because the two directions are evaluated under different test-set conditions, the conclusion that 'models trained on social media struggle to generalize to minority-specific contexts, and vice-versa' is stronger than the experiments support. At minimum, the discussion should restate the claim conditionally on the synthetic nature of the minority-focused test sets, or the authors should add the missing real-data evaluation.
minor comments (6)
- [§3.2] The text says the expansion stage covers nine target communities but then enumerates only eight (Black people, Jews, Muslims, Indigenous Brazilians, women, LGBTQIA+ individuals, elderly people, and people with disabilities). The ninth group, 'immigrants,' appears in Table 1 and Section 3.3 but is missing from the list.
- [§1] There are several typos and grammatical slips in the introduction, including 'only only toxic or non-toxic annotations' and 'current existing datasets suffers from a severe domain confinement.'
- [Table 1] The grand total is printed as '53.274' with a period; it should be '53,274' for consistency with the rest of the paper.
- [§3.4 and §3.7] The relation between the 'final minority-targeted corpus of 50,074 sentences' in Section 3.4 and the '53,274 synthetic sentences' in Section 3.7 is not explicit; clarify that 53,274 = 50,074 + 3,200 neutral samples.
- [Abstract and title] The arXiv title uses 'ToxSyn-PT' but the abstract and body use 'ToxSyn'; this naming inconsistency should be resolved in the final version.
- [§9, Ethics Statement] The sentence 'applications that intend to use ToxSyn should be validated by a under multidisciplinary team' contains a grammatical error ('a under') that should be corrected.
Circularity Check
No significant circularity: the dataset construction and cross-domain experiments are empirical and externally grounded, with no fitted parameter or self-citation chain substituting for the claimed results.
full rationale
This paper does not derive any quantity from an input in a way that makes the output equivalent to the input by construction. The central claim of a catastrophic, mutual generalization failure is an empirical observation obtained by fine-tuning BERTimbau on one corpus and testing on held-out portions of external corpora (Section 4.3, Table 4). The ToxSyn test set is human-annotated by a red team blind to the machine labels (Section 3.6), so the reported in-domain performance is not circular with respect to the generation pipeline. The cross-domain results use external benchmarks ToLD-BR, HateBR, OLID-BR, and a translated version of ToxiGen; none of these are fitted parameters in the model, and the paper reports actual recall and F1 scores. The main validity concern—that ToxSyn and translated ToxiGen are both LLM-generated and may share synthetic artifacts—is a limitation explicitly acknowledged in Section 8 ('repetitive phrasing and distributional artifacts') and Section 4.2 ('unnaturalized phrasing and culturally mismatched slurs'), but it is a threat to external validity, not circular reasoning. There is no self-citation used as load-bearing evidence, no uniqueness theorem invoked from the authors' prior work, and no ansatz smuggled in through citation. The paper's claims are therefore self-contained against external benchmarks, and the derivation chain is not circular.
Assumptions & free parameters
free parameters (4)
- Seed dataset size =
40 toxic + 40 non-toxic per group (160 total)
- Generation template catalog size =
26 templates, then 28 in enrichment; 20 style-transformation prompts for neutral samples
- Neutral sample integration size =
3,200 sampled from 6,000 candidates
- Group-specific counterexample sampling =
3,000 per group: 1,000 same-group non-toxic, 1,000 other-group toxic, 1,000 other-group non-toxic/neutral
assumptions (4)
- domain assumption LLM-generated synthetic sentences are a valid proxy for real-world minority-targeted hate speech in Portuguese.
- domain assumption The GPT-4-translated human-annotated ToxiGen subset is a valid Portuguese benchmark for minority-targeted toxicity.
- domain assumption Human annotations on the 5,208-item test set are a reliable gold standard.
- domain assumption Merging categories across Portuguese datasets preserves the minority-target toxicity construct.
Cite this review
Pith. "Pith review of ToxSyn-PT: A Synthetic Fine-Grained Dataset of Minority-Targeted Toxic Language in Portuguese." pith.science (2026). https://pith.science/paper/DNIDRGUO
@misc{pith2026250610245,
author = {Pith},
title = {Pith review of: ToxSyn-PT: A Synthetic Fine-Grained Dataset of Minority-Targeted Toxic Language in Portuguese},
year = {2026},
howpublished = {\url{https://pith.science/paper/DNIDRGUO}},
note = {Machine review of arXiv:2506.10245}
}
read the original abstract
The development of robust hate speech detection systems remains limited by the lack of large-scale, fine-grained training data, especially for languages beyond English. Existing corpora typically rely on simplistic toxic and non-toxic labels, and the few that capture hate directed at specific minority groups lack the positive counterexamples required to distinguish genuine hate from mere discussion. In this work, we introduce ToxSyn-PT, the first Portuguese large-scale corpus explicitly designed for multi-label hate speech detection across nine protected minority groups, including the non-toxic counterexamples absent in all other public datasets. Generated via a controllable four-stage pipeline, ToxSyn contains discourse-type annotations to capture rhetorical strategies of toxic/non-toxic language, such as sarcasm, dehumanization, and cultural appreciation. Our experiments reveal a catastrophic, mutual generalization failure compared to existing datasets from social-media domains: models trained on social media struggle to generalize to minority-specific contexts, and vice-versa. This finding indicates they are distinct tasks and exposes summary metrics like Macro F1 can be unreliable indicators of true model behavior, as they completely mask model failure. We publicly release ToxSyn on HuggingFace to support reproducible research on synthetic data generation and benchmark progress in hate-speech detection for low- and mid-resource languages.
Figures
Forward citations
Cited by 1 Pith paper
-
Safety Is Not Universal: The Selective Safety Trap in LLM Alignment
Safety alignment in LLMs is not uniform but forms a demographic hierarchy, with defense rates varying by up to 42% across groups; a new benchmark and DPO method demonstrate transferable safety.
Reference graph
Works this paper leans on
-
[1]
However, progress is limited by severe limitations in available training corpora
Introduction The task of identifying and mitigating online hate speech is a critical challenge for building safe and inclusive digital spaces, from social media plat- forms to nascent virtual reality ecosystems (Albladi et al., 2025; Weerasinghe et al., 2025). However, progress is limited by severe limitations in available training corpora. Most datasets ...
arXiv 2025
-
[2]
WeintroduceToxSyn,thefirstPortuguesepub- liclyavailablecorpusdesignedtosupporthate- speech classification across multiple minor- ity targets, comprising over 50K synthetic in- stances annotated with toxicity, target group, and discourse-type labels
-
[3]
We present a controllable LLM-based data generation pipeline that allows for balancing class distributions, injecting low-frequency ex- pressions, and applying safety constraints
-
[4]
We demonstrate that toxicity detection in Por- tuguese is strongly domain-dependent, with model performance degrading sharply when applied to out-of-distribution contexts
-
[5]
Related Works Synthetic corpora have been explored as a means to augment limited hate speech datasets. ToxiGen (Hartvigsen et al., 2022) combines a classifier-in- the-loopframeworkwith GPT-3(Brown et al.,2020) to adversarially generate over 250,000 examples across 13 demographic targets, yielding measur- able improvements in classification performance. To...
work page 2022
-
[6]
Dual paraphrase augmentation triples corpus size and stylistic variety
Seed Dataset 2) Prompt-Based Expansion 3) Paraphrase Augmentation 4) Enrichment Few ‑ shot prompting creates minority ‑ focused sentences based on predefined strategies. Dual paraphrase augmentation triples corpus size and stylistic variety. Enrich with implicit ‑ hate variants using the generated dataset. Figure 1: Overview of ToxSyn generation pipeline,...
-
[7]
ToxSyn This section describes the construction of ToxSyn, a large-scale dataset designed to provide compre- hensive coverage of minority-targeted hate speech in Portuguese. Our approach emphasizes con- trolled representation of protected groups and lin- guistic diversity across both toxic and non-toxic expressions. The dataset is produced through a struct...
-
[8]
Limitations While ToxSyn substantially advances Portuguese hate-speech resources, it carries several impor- tant caveats. First, each sentence is annotated with a single target group, precluding the analysis of intersectional or multi-target cases; overlapping abuses (e.g., simultaneously sexist and racist lan- guage) thus remain unmodeled. Second, the co...
Show all 20 references
-
[9]
The results of this validation strongly affirm the high fidelity of our synthetic data
to directly quantify the quality of our synthetic generation pipeline. The results of this validation strongly affirm the high fidelity of our synthetic data. The human an- notators modified around 7% of the primary toxi- city labels and fewer than 1% of the specific mi- norit...
-
[10]
watermelon
Experiments We evaluate the effectiveness of the ToxSyn dataset under two classification settings: (1) multi- domain classification, which determines whether a given text is toxic or non-toxic in both general and minority targeted benchmarks, and (2) group- specific classifica...
2022
-
[11]
Discussion Our experiments reveal a profound domain de- pendency in Portuguese toxicity detection. The cross-domain evaluation demonstrates a mutual and catastrophic generalization failure: models Group Macro-Recall Support PT Merged Black 0.63 230 Women 0.63 961 LGBTQIA+ 0.64...
-
[12]
Conclusion In this paper, we introduced ToxSyn, a large-scale, fine-grained synthetic dataset that fundamentally addresses a critical resource gap in Portuguese. It is the first Portuguese corpus to enable the fine- grained classification of hate against specific mi- nority gr...
-
[13]
First, our four-stage gen- eration pipeline serves as a generalizable blueprint
Future Work Our work on ToxSyn opens several promising av- enues for future research. First, our four-stage gen- eration pipeline serves as a generalizable blueprint. A clear next step is to adapt this methodology to otherlow-resourcelanguagesthatfacesimilardata scarcity, enab...
-
[15]
Ethics Statement The use of generative models to create synthetic toxic data must be approached with great caution to prevent harmful applications, such as deliber- atelyoffendingminoritygroupsortraininglanguage models to produce hate speech. Nevertheless, re- sponsible use of...
-
[16]
Bibliographical References Josh Achiam, Steven Adler, Sandhini Agar- wal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Al- tenschmidt,SamAltman,ShyamalAnadkat,etal
-
[17]
Aish Albladi, Minarul Islam, Amit Das, Maryam Big- onah, Zheng Zhang, Fatemeh Jamshidi, Mostafa Rahgouy, Nilanjana Raychawdhary, Daniela Marghitu, and Cheryl Seals
Gpt-4 technical report.arXiv preprint arXiv:2303.08774. Aish Albladi, Minarul Islam, Amit Das, Maryam Big- onah, Zheng Zhang, Fatemeh Jamshidi, Mostafa Rahgouy, Nilanjana Raychawdhary, Daniela Marghitu, and Cheryl Seals. 2025. Hate speech detection using large language models:...
2025 arXiv
-
[20]
2017.Offensive comments in the brazilian web: a dataset and baseline results
Language Resource References De Pelle, Rogers Prates and Moreira, Viviane P. 2017.Offensive comments in the brazilian web: a dataset and baseline results. SBC. Fortuna,PaulaandRochadaSilva,JoãoandSoler- Company, Juan and Wanner, Leo and Nunes, Sérgio. 2019.A Hierarchically-Lab...
2017
-
[2023]
immigrant from *
in a few-shot prompting configuration to gen- erate new samples conditioned on minority identity. The model is prompted with representative seed examples and instructed to produce linguistically diverse toxic and non-toxic statements for one of nine target communities: Black p...
2023
-
[2024]
InProceedings of the 16th International Con- ference on Computational Processing of Por- tuguese, pages 108–116
Toxic speech detection in portuguese: A comparative study of large language models. InProceedings of the 16th International Con- ference on Computational Processing of Por- tuguese, pages 108–116. Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and...
2022 arXiv
-
[2025]
Beyond mute and block: adoption and effectiveness of safety tools in social vr, from ubiquitous harassment to social sculpting.IEEE Transactions on Visualization and Computer Graphics
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.