Data & software
Collections, models and code
Most of our papers come with something others can reuse. First the resources we maintain, then everything we have published on GitHub and HuggingFace.
- Shared task
eRisk: Early Risk Prediction on the Internet
The CLEF evaluation lab on early risk detection, run every year since 2017. Recent editions cover symptom ranking, contextual and conversational approaches to depression and ADHD. The 2027 edition is part of CLEF 2027.
- Dataset
MetaHate
A meta-collection of 1,226,202 social media posts labelled as hate or non-hate, unified from public hate speech datasets so that models can be trained and compared on a common ground.
Access. Gated, for research use under the terms of the source datasets. A 100-post sample is open.
- Dataset
ReDSM5
1,484 long Reddit posts annotated sentence by sentence by a licensed psychologist for the nine DSM-5 depression symptoms, each label with a short clinical rationale.
Access. Gated: a 25-post sample is public; the full set is available on request.
- Models
eRisk 2026 personas and adapters
LLM-based patient personas and adapters released for the conversational tasks of eRisk 2026, downloaded more than 1,000 times.
- Dataset
Mute Cods
A multilingual Telegram dataset with benchmark models for conspiracy theory detection, built within the HYBRIDS network.
- Dataset
IAT2BAS
A dataset released in 2026 with the Fondazione Bruno Kessler and CSIC.
- Software
LabChain
A Python framework for reproducible and modular scientific experiments, developed with the Universidade de Santiago de Compostela.
Everything we release 52
12 datasets and 19 models on HuggingFace, 21 repositories on GitHub. Gated items are approved by the authors on request.
Datasets
- Hate speech ICWSM 2024
MetaHate
A meta-collection of 36 hate speech datasets from social media, unified under one binary label. Version 2.0 refines cross-dataset deduplication. MetaHateES is the Spanish counterpart.
- Mental health CIKM 2025
ReDSM5
Reddit posts annotated sentence by sentence for the nine DSM-5 symptoms of a major depressive episode, each label with an expert rationale. A paraphrased 25-post sample needs no agreement.
- Mental health
eRisk Depression Severity Personas
Personas of 116 Reddit users with self-reported depression and their BDI-II severity, for persona-conditioned user simulation (Eval4MHSim). A paraphrased three-user sample is open.
- Hate speech
Induced Hate Speech Classification Guidelines
Classification guidelines induced by an LLM from labelled examples of the Measuring Hate Speech corpus. They quote hateful posts verbatim, so access is approved manually.
- Crawling
ProdCrawl
A replayable test collection for focused crawling: static mirrors of 621 company websites with a relevance judgment for every page for "is this a product page?", plus the fine-tuned anchor-text encoder.
- RAG & search
LoreHopQA
Multi-hop questions set in 61 invented worlds, so the retrieved context is the only source of the answer. Gated so that no evaluated model sees it in training; evaluation and publication are unrestricted.
- Galician arXiv 2023
Galician instruction data
Galician translations of the Stanford Alpaca instructions (52K) and of ShareGPT conversations, used to train Cabuxa.
Models
- Hate speech
MetaHate classifiers
BERT and multilingual BERT fine-tuned on MetaHate for binary hate speech detection in English, Spanish, Portuguese and Galician.
- Hate speech ECIR 2025
Distil-MetaHate
Llama 3 8B distilled with chain-of-thought to classify hate speech and explain why.
- Hate speech NAACL 2025
Stop-Hate
Mistral 7B and Llama 2 13B fine-tuned so that conversational models do not engage with or amplify hate speech.
- Hate speech ICWSM 2026
Geographic debias tuning
Llama 3.1 8B, Mistral Nemo and Phi-4 mini adapters debias-tuned so that hate speech decisions stay consistent when a country persona is added, with and without language context.
- Mistral-Nemo-Instruct-2407-Geographic-Debias-Tuning
- Mistral-Nemo-Instruct-2407-Geographic-Debias-Tuning-Lang
- Phi-4-mini-instruct-Geographic-Debias-Tuning-Lang
- Phi-4-mini-instruct-Geographic-Debias-Tuning
- Llama-3.1-8B-Instruct-Geographic-Debias-Tuning
- Llama-3.1-8B-Instruct-Geographic-Debias-Tuning-Lang
- Galician arXiv 2023
Cabuxa: LLMs for Galician
Cabuxa-7B, a LLaMA instruction model for Galician, and Cabuxa 2.0, Llama 3.1 8B Instruct with continued pre-training on the CorpusNós corpus, also as a 4-bit GPTQ build.
- Mental health ECIR 2025
MindWell
Adapter for Vicuna 7B behind MindWell, a conversational agent that supports professional depression screening on social media.
- Crawling
ProdCrawl anchor-text encoder
MPNet fine-tuned to score anchor text for links that lead to product pages, used by the ProdCrawl focused crawler.
Code
- prodcrawl
Focused crawler with BFS, DFS, classifier and retrieval-based anchor scoring, and the scripts to serve the ProdCrawl mirrors.
- lorehop-qa
Validation experiments for LoreHopQA: closed-book answering, ablated evidence and complete context.
- eval4sim
Eval4Sim, an evaluation framework for persona simulation with LLMs.
- eval4mhsim
Eval4MHSim, a persona collection and evaluation framework for simulating social media users with mental health conditions.
- llm-prf
LLM-assisted pseudo-relevance feedback that filters noisy feedback documents before query expansion.
- erisk-qrels-expansion
Champion runs and LLM-assisted annotation to extend the shallow relevance judgments of eRisk collections.
- recsysipp
LLM relevance judgments for evaluating recommender systems on MovieLens: judgment quality, augmented pools and popularity bias.
- grpo-dsm5
Group Relative Policy Optimization to make LLM reasoning about DSM-5 depression symptoms clinically reliable.
- rcl
Hybrid pooling with LLMs through relevance context learning.
- hate-eval-agreement
Whether LLMs can evaluate hate speech classifiers even where they annotate poorly.
- geographic-bias
Geographic bias in LLM hate speech detection, and debias tuning.
- multiple-testing
Statistical testing for comparing many retrieval systems at once.
- Lost-in-the-Evidence-in-RAG-Virtual-Appendix
Virtual appendix for a reproducibility study of document position and context size effects in RAG.
- bayesian-inference
Bayesian inference with Stan for statistical comparisons between retrieval systems.
- SIQSE
Simulates users choosing query suggestions with LLMs, to evaluate suggestions beyond the top positions.
- depression-prompt-induction
Symptom Induction: a teacher LLM distils labelled examples into interpretable guidelines for spotting evidence of depression symptoms.
- ColBERTKP
ColBERT adapted to keyphrase queries.
- distil-metahate
Distilling LLMs into efficient, explainable hate speech classifiers.
- decoding-hate
How language models react to hate speech, and how to stop them engaging with it.
- metahate
Scripts that build MetaHate from its 36 source datasets.
- RAKS
Retrieval-augmented generation for suggesting keyphrase synonyms.