SINAI/ALIA-es-discriminative-hate-speech

Dataset Introduction

The ALIA Spanish Discriminative Hate Speech Corpus is a large-scale Spanish dataset for hate-speech detection built from curated social-media comments and automatically annotated using a multi-expert LLM pipeline with Fusion of Experts (FoE)[1].

The release contains:

• 228,708 instances
• Spanish comments from YouTube and TikTok
• Per-expert predictions and explanations from three LLM experts
• Final fused outputs (foe_class, foe_score) for discriminative hate-speech classification
• Human validation subset of 3,000 comments annotated by two native Spanish speakers via Prolific to assess LLM and fusion agreement

The corpus is intended for research on Spanish hate-speech detection, model disagreement analysis, explanation-aware moderation, and confidence-based evaluation. All documentation and source code for the collection, curation, annotation, fusion, and human validation processes are available in the ALIA-UJA GitHub repository.

Dataset Details

Dataset Description

This dataset is the final fused artifact of a three-stage pipeline:

1. Raw social-media collection (YouTube Data API v3 and TikTok Research API).
2. Curation and filtering of Spanish comments (language filtering, spam heuristics, anonymization, optional near-duplicate removal).
3. Automatic annotation by three prompted LLM experts and FoE fusion.

Each record preserves both the expert-level outputs and the fused decision. This allows users to work with a standard binary hate label while also analyzing disagreement and rationale patterns across experts.

Dataset Sources

• Primary social sources: YouTube comments and TikTok comments.
• Collection interfaces: Official platform APIs.
• Processing context: Curated and merged Spanish comment corpus with source metadata.

Uses

The primary purpose of this corpus is to support training and evaluation of Spanish hate-speech detection systems, with applications in:

• Binary hate-speech classification.
• Confidence calibration using continuous fused scores.
• LLM annotation reliability and disagreement studies.
• Explanation-based moderation analysis.
• Comparative evaluation of expert-only vs fused predictions.

Dataset Structure

Data Instances

Each instance follows this structure:

{
  "id": "record_199077",
  "source": "youtube",
  "text": "Moros fuera TVE sicópatas iros a marruecos",
  "foe_class": 1,
  "foe_score": 0.9392520189285278,
  "apertus_it_8B_class": 1,
  "apertus_it_8B_explanation": "El comentario contiene lenguaje degradante hacia colectivos vulnerables (moros) y utiliza un término despectivo (sicópatas) que deshumaniza a las personas de origen marroquí. Además, incita a la discriminación y exclusión al pedir que se vayan a Marruecos.",
  "gemma_4_31B_it_class": 1,
  "gemma_4_31B_it_explanation": "El comentario constituye discurso de odio ya que utiliza el término 'moros' de forma despectiva y exige la expulsión de un grupo basado en su origen nacional o etnia ('fuera', 'iros a marruecos'). Esto cumple con los criterios de llamar explícitamente a la exclusión y segregación de un grupo protegido, promoviendo la xenofobia.",
  "minimax_m2.5_class": 1,
  "minimax_m2.5_explanation": "El comentario 'Moros fuera TVE sicópatas iros a marruecos' contiene discurso de odio porque utiliza el término despectivo 'moros' para referirse a personas de origen árabe o norteafricano, lo cual es un insulto étnico. La expresión 'fuera' implica una llamada a la expulsión o exclusión de este colectivo, lo cual constituye incitación a la discriminación y segregación. Además, el insulto 'sicópatas' directed toward TVE (la televisión pública española) y la orden 'iros a marruecos' refuerzan el mensaje de rechazo hacia personas de origen marroquí o magrebí. El comentario combina un insulto étnico con una llamada explícita a la exclusión, cumpliendo los criterios de discurso de odio."
}

Data Fields

Field	Type	Description
id	string	Unique record identifier in the final release.
source	string	Platform source (youtube or tiktok).
text	string	Spanish comment text.
foe_class	int	Fused binary label (0 = no-hate, 1 = hate).
foe_score	float	Fused positive-class confidence score.
apertus_it_8B_class	int	Expert binary prediction (0 or 1) from apertus_it_8B.
apertus_it_8B_explanation	string	Spanish generated rationale from apertus_it_8B.
gemma_4_31B_it_class	int	Expert binary prediction (0 or 1) from gemma_4_31B_it.
gemma_4_31B_it_explanation	string	Spanish generated rationale from gemma_4_31B_it.
minimax_m2.5_class	int	Expert binary prediction (0 or 1) from minimax_m2.5.
minimax_m2.5_explanation	string	Spanish generated rationale from minimax_m2.5.

Data Splits

The released JSONL is provided as a single consolidated split.

Distribution of final fused labels:

Label	Count	Percentage
foe_class = 0 (no-hate)	167,774	73.36%
foe_class = 1 (hate)	60,934	26.64%
TOTAL	228,708	100.00%

Distribution by source:

Source	Count	Percentage
youtube	201,612	88.15%
tiktok	27,096	11.85%
TOTAL	228,708	100.00%

Expert-level agreement summary (common IDs):

Metric	Value
Fleiss' Kappa (3 experts)	0.49827
Majority-vote hate	25.97%
Majority-vote no-hate	74.03%

Example Usage

from datasets import load_dataset

dataset = load_dataset(
  "SINAI/ALIA-es-discriminative-hate-speech",
  split="train",
)

print(dataset[0]["id"])
print(dataset[0]["source"])
print(dataset[0]["foe_class"], dataset[0]["foe_score"])

Streaming mode:

from datasets import load_dataset

stream_ds = load_dataset(
  "SINAI/ALIA-es-discriminative-hate-speech",
  split="train",
  streaming=True,
)

for i, ex in enumerate(stream_ds):
  print(ex["id"], ex["foe_class"], ex["source"])
  if i >= 2:
    break

Dataset Creation

Curation Rationale

This corpus was created to provide a large Spanish discriminative resource for hate-speech detection in social media. The project objective is not to provide a human-ground-truth moderation dataset as these annotations are synthetic. Instead, the release aims to collect a broad corpus of social-media comments annotated by multiple LLM experts whose labels and explanations may differ (analogous to human annotators), providing a synthetic resource for analysis and evaluation with human expert annotations.

Source Data Collection

The data originates from YouTube and TikTok comments collected through official APIs and processed into a curated Spanish corpus.

The collection strategy emphasizes the kinds of hate narratives that are most visible in Spanish social-media discourse. The corpus is especially useful for analyzing comments related to:

• Immigration and xenophobia.
• Public safety and insecurity narratives.
• Political polarization and partisan antagonism.
• Social conflict associated with high-salience local debates and events.

For a detailed description of the collection process, including query strategies, collection justifications, and source metadata, see Data Collection Documentation or the source code repository.

Data Collection and Processing

The processing workflow includes collection, curation and deduplication. The curation pipeline is implemented using the datatrove toolchain and custom LambdaFilters for filtering and normalization. The main stages are ordered as follows:

1. Emoji-only removal (LambdaFilter): drop comments that are only emojis or whitespace.
2. Language detection (LanguageFilter using fastText): requires language_threshold >= 0.7 (default: 0.7) to retain Spanish comments.
3. Spam heuristics (LambdaFilter): default thresholds include max_urls=1, max_mentions=5, max_repeated_seq=15 (identical chars), and max_nonalpha_ratio=0.7.
4. Anonymization (LambdaFilter): replace @... mentions with the canonical @usuario.
5. MinHash deduplication: character 5-grams MinHash signatures, bucketing and clustering; note the deduplication step requires sequential processing (world_size=1) and removes near-duplicates while keeping one representative per cluster.

For a detailed, step-by-step description including intermediate counts, MinHash configuration and stage timings, see Curation Data Process Documentation or the source code repository.

Annotations

This resource does not include human-curated final labels. All annotations are produced automatically by prompted LLM experts and then fused by a learned FoE [1] artifact. See the full annotation and fusion workflow in Annotation Data Process Documentation or the source code repository.

Annotation pipeline:

• Per-model annotation is performed with local vLLM checkpoints via an outlines offline wrapper. The annotation runner loads a prompt template and enforces a structured output schema (HateClass) implemented with Pydantic. The schema requires two fields: is_hate (boolean) and explanation (Spanish string).
• Generation parsing uses json_repair to tolerate minor formatting issues; the pipeline appends valid responses to per-model JSONL files (one JSONL per model alias) which act as the source-of-truth for subsequent passes.
• Retries are performed for transient failures; permanently unresolved items are marked with __EXHAUSTED__ in the explanation field while keeping is_hate = null.

Models used for annotation (per-model outputs are merged later):

• swiss-ai/Apertus-8B-Instruct-2509 (alias apertus_it_8B)
• google/gemma-4-31B-it (alias gemma_4_31B_it)
• MiniMaxAI/MiniMax-M2.5 (alias minimax_m2.5)

Fusion stage (high level):

• Per-model JSONL outputs are read by small MLP architecture, which builds feature vectors concatenating: binary annotations (per model), comment embeddings, and explanation embeddings (both computed by sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2).
• The selected FoE artifact produces the final fused outputs written into the consolidated JSONL: foe_class (binary) and foe_score (positive-class probability).

FoE [1] configuration was selected by following Fusion of Experts (FoE) Experimentation Methodology for Spanish Hate Speech Detection, which includes a systematic evaluation of multiple fusion configurations. The best-performing configuration on the pooled validation sets (and selected for the final release) is:

• fusion head: sentiment_mlp
• embedding method: sentence_transformer
• PCA components: 128
• hidden dim: 1024
• dropout: 0.5
• batch size: 64
• random state: 42

Persisted model metrics:

Split	Accuracy	F1-macro	Kappa
Pooled validation	0.8230	0.8194	0.6389
Pooled test	0.8179	0.8112	0.6225

Human Validation (Prolific)

To assess the reliability of the automated annotation pipeline and the Fusion of Experts (FoE) classifier, a validation study was conducted with human annotators through the Prolific platform.

• Participant Selection: Two native Spanish speakers from Spain (anonymized as Annotator 1 and Annotator 2) were selected under strict demographic filters (Spanish nationality, country of birth, current residence, and first/primary language).
• Sampling Strategy: A representative sample of 3,000 comments was annotated, balanced across platforms (1,500 YouTube and 1,500 TikTok). The sample was selected using stratified diversity sampling to preserve the proportions of positive and negative classes predicted by the FoE model, and applying Maximal Marginal Relevance (MMR, $\lambda = 0.7$) over TF-IDF comment vectors to maximize semantic diversity.
• Agreement Analysis:
  • Inter-Annotator Agreement: Cohen's Kappa of 0.510 (Accuracy: 0.892, F1-macro: 0.753), indicating moderate-to-good agreement.
  • Human vs. FoE Classifier:
    • Annotator 1 vs FoE: Cohen's Kappa of 0.485 (Accuracy: 0.866, F1-macro: 0.737)
    • Annotator 2 vs FoE: Cohen's Kappa of 0.448 (Accuracy: 0.837, F1-macro: 0.723)
  • Global Agreement (Annotators + FoE): Fleiss' Kappa of 0.474 (moderate agreement).

For full details on the participant profile, TF-IDF candidate prefiltering, MMR formulation, and detailed agreement matrices across all expert LLM models, see the Human Validation and Agreement (Prolific) Section in the repository.

Personal and Sensitive Information

The source is user-generated social-media text and may contain sensitive, abusive, or harmful language. Mentions are anonymized during curation, but users should apply additional safeguards depending on downstream use and legal requirements.

Considerations for Using the Data

Social Impact of Dataset

This dataset can improve Spanish hate-speech moderation research and help build safer content analysis systems. At the same time, it contains potentially harmful expressions and should be handled with responsible access, annotation transparency, and clear downstream risk controls.

Discussion of Biases

Potential biases include:

• Platform bias (YouTube-heavy distribution).
• Temporal/contextual bias tied to collection windows.
• Model-induced bias from prompted LLM experts.
• Fusion-model bias inherited from training datasets and feature engineering choices.
• Possible underrepresentation of subtle hate patterns not consistently captured by all experts.

Other Known Limitations

• Not a human-ground-truth gold label release for this final corpus.
• Explanations are generated model rationales and may include hallucinations.
• Single consolidated split; no official benchmark split is included.
• Social-media content and platform policy constraints may limit certain redistribution or deployment scenarios.
• Binary labels can compress nuanced, context-dependent harmful discourse.

Citation

@misc{ALIA-es-discriminative-hate-speech,
  title={ALIA Spanish Discriminative Hate Speech Corpus},
  author={SINAI Research Group},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/SINAI/ALIA-es-discriminative-hate-speech}}
}

Please also cite:

@inproceedings{ICLR2024_c5169260,
 author = {Wang, Hongyi and Maia Polo, Felipe and Sun, Yuekai and Kundu, Souvik and Xing, Eric  and Yurochkin, Mikhail},
 booktitle = {International Conference on Learning Representations},
 editor = {B. Kim and Y. Yue and S. Chaudhuri and K. Fragkiadaki and M. Khan and Y. Sun},
 pages = {45284--45306},
 title = {Fusing Models with Complementary Expertise},
 url = {https://proceedings.iclr.cc/paper_files/paper/2024/file/c5169260ef32d1bd3597c14d8c89b034-Paper-Conference.pdf},
 volume = {2024},
 year = {2024}
}

---

Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU - NextGenerationEU within the framework of the project ALIA.

Acknowledgments

This dataset has been generated thanks to SCAYLE (Centro de Supercomputación de Castilla y León), which provided computational resources on its CALENDULA supercomputing cluster.

References

[1] Wang, Hongyi, Felipe Maia Polo, Yuekai Sun, Souvik Kundu, Eric Xing, and Mikhail Yurochkin. "Fusing Models with Complementary Expertise." International Conference on Learning Representations (ICLR), 2024. https://proceedings.iclr.cc/paper_files/paper/2024/file/c5169260ef32d1bd3597c14d8c89b034-Paper-Conference.pdf

---

Contact: ALIA Project - SINAI Research Group - Universidad de Jaén

More Information: SINAI Research Group | ALIA-UJA Project