We present two competitive solutions in the Te Ahorré Un Click (TA1C) competition for (i) clickbait detection in Spanish tweets and (ii) clickbait spoiling given a headline and an article body. For the detection task, multiple BERT encoders were benchmarked, obtaining a blind-test macro-F1 of 0.80480, ranking 4th on the blind test set. On the spoiler generation task, we designed a pipeline and tested multiple axes of variation. These included the use of information retrieval (top- chunks) versus no retrieval (using the full article on the prompt), prompt language (Spanish vs. English), number of few-shot demonstrations (0-3), and whether to inject definitions of clickbait and spoiler. Supervised fine-tuning and multiple inference generation, ranked by a separate “judge” model, provided a significant upgrade over the Gemini-1.5-flash-002 model. The final spoiling pipeline achieved a BLEU metric of 0.43589 on the blind test set, ranking 1st in spoiler generation.
Clickbait is a widespread practice in online journalism: headlines or teasers deliberately omit a key
piece of information in order to trigger curiosity and force the reader to click. Curiosity arises when
people notice a gap between what they know and what they want to know; clickbait exploits that gap
to capture trafic. The Te Ahorré Un Click (TA1C) shared task [
In this paper, Section 2 introduces the two TA1C subtasks and the corpora on which they are evaluated. Section 3 details our methodology for both clickbait detection and spoiler generation, including model architectures, prompt design and experimental protocol. Section 4 reports and analyses the results obtained in each task. Finally, Section 5 summarises the main findings and sketches directions for future work.
The corpus comprises 4200 Spanish tweets posted between October 2020 and October 2021 by 18 major
media outlets, spanning twelve national varieties [
A manually curated subset of 500 tweets (300 train, 100 dev and 100 test examples) extends the detection corpus with human-written spoilers that close the information gap. The answers cover factoids, summaries, lists and “no-answer” cases. Table 2 presents an abbreviated example. The full article text is stored as JSON and omitted here for space. 1355225506040438784 2021-01-29 Cómo inversores aficionados se enfrentaron a . . .
La clave son las “ventas en corto”, en las que un fondo apuesta a que el precio caerá.
Our detection pipeline is simple. We benchmarked a small pool of Spanish (and one multilingual)
BERT-type encoders. The evaluated models were PlanTL-GOB-ES/roberta-large-bne[
Figure 1 presents a detailed overview of our spoiler generation pipeline. The process begins with
the user providing the input pair: a tweet and its linked article with optional metadata (country and
outlet) that we choose to include in the article body if requested. This input is first preprocessed (e.g.
truncation, format cleaning), after which the article is then split into key chunks of sentences each,
using a semantic model, allowing the possibility of using the whole body in the prompt. Then, the
system branches depending on whether KATE-style[
If KATE-style retrieval is active, the title is first embedded with all-MiniLM-L6-v2[
Once the examples are selected, the system builds a complete prompt by concatenating a task-specific
system message based in similar works[
In the second phase, we also introduced a two-stage generation and correction pipeline, and a postprocessing final step. At first, the chosen writer LLM produces distinct candidate spoilers. This generates M candidates that can present variations in their structure. Then, a dedicated judge LLM reviews all M outputs, choosing the best one according to a tiny prompt we wrote for this final step (Figure 3). We propose this fix as external to the final model, and present 4 final models for the evaluation phase, varying only this parameter. Finally, we use a postprocessing step to remove manually detected prompt-related inaccuracies generated by the models (e.g. "SALIDA ESPERADA:" message at the start).
--- SYSTEM --Eres un experto en redactar spoilers de clickbait para arruinar las noticias clickbait en español. El clickbait es un método para generar titulares llamativos, especialmente en línea, que omite deliberadamente parte de la información con el objetivo de generar curiosidad, al crear una brecha de información, atrayendo así la atención de los lectores y logrando que hagan clic.
Un spoiler de clickbait es una corta respuesta objetiva que acaba con la brecha de información que el Tweet siembra y que, en el artículo, se resuelve.
La principal funcionalidad del spoiler es que el lector ya no necesite abrir la noticia. Tu tarea es leer el titular y el artículo, y escribir una única oración en español que revele claramente lo que el tweet insinúa. Busca en el artículo la frase o dato que responda a la duda planteada, preferiblemente citando el texto original entre comillas si se trata de una declaración. Sé conciso: redacta una sola oración que contenga la idea central, sin repetir nombres, fechas, hashtags ni frases o entidades ya presentes en el tweet.
Ten presente que pueden existir los siguientes casos particulares: 1. Que no exista respuesta en el conjunto. En ese caso debes responder SOLAMENTE (sin comillas): No hay respuesta 2. Que la respuesta sea la nota completa. En ese caso debes responder SOLAMENTE (sin comillas): La respuesta es la nota completa --- EXAMPLES --Ejemplos (titular, artículo, frase reveladora): TWEET: ...
NOTICIA ORIGINAL: ...
RAZONAMIENTO: Paso 1: ¿Cuál es/son la(s) parte(s) oculta(s) a desvelar en el tuit?: ...
Paso 2: Por lo tanto, ¿qué pregunta(s) debemos responder?: ...
Paso 3: ¿Aparece(n) la(s) respuesta(s) en la noticia? ¿Dónde exactamente (cita breve o localiza el párrafo)?: ...
Paso 4: ¿Cómo resumirías esa(s) respuesta(s) sin repetir datos del tuit?: ...
Paso 5: Por lo tanto, la respuesta es: ...
SALIDA ESPERADA (debes escribir lo siguiente): ... --- TASK --TWEET: {tweet} NOTICIA ORIGINAL: {article}
SALIDA ESPERADA:
Elige el mejor spoiler entre los 3 candidatos. Criterio: Responde la incógnita del tuit de forma correcta, concisa y sin repetir literalmente el tuit. Tweet: ¿Sabes qué jugador fue transferido por cifra récord? Artículo (referencia): El club confirmó que Juan Pérez fue traspasado al Real FC por 50 millones de euros, estableciendo un nuevo récord en la liga...
Candidatos: 1. Fue Juan Pérez, vendido por 50 millones de euros. 2. El club no ha confirmado ninguna venta. 3. Se trata de Luis Gómez.
Respuesta:
In the detection task, we evaluated a range of Spanish BERT variants in the clickbait detection task using a 5-fold cross-validation protocol with class-weighted loss, using both train and val datasets, training each model for three epochs at a learning rate of 2 × 10− 5, and a batch size of 16.
For the spoiling task, the evaluation was split into three phases. At the very start of the competition, we set aside 60 training instances and benchmarked five candidates’ LLMs. Then, we chose the best local model and then began refining the prompt using this model. Prompt configuration included evaluation for article information retrieval, few-shot example building variants, clickbait and spoiler definitions, and prompt language. When prompt construction was done, multiple LLM temperatures were evaluated by a sweep, and supervised fine-tuning was also explored.
The best base model PlanTL-GOB-ES/roberta-large-bne[
We then refined the prompting strategy using Gemma-3 12B. Initially, we compared three article retrieval methods: using the full article versus using smaller fragments (five fragments of two sentences 1https://huggingface.co/dcere/ta1c-Clickbait-Detector-es-large or two fragments of two sentences each). As shown in 6, providing the model with the complete article consistently outperformed fragmented versions, which frequently omitted critical information, causing the model to mistakenly predict the fallback answer "No hay respuesta".
Next, we examined few-shot prompting strategies. The KATE-style example retrieval—MiniLM
embeddings [
We also assessed the impact of including explicit definitions of clickbait and spoiler. Although adding these definitions did not noticeably alter performance metrics, we included them in the final pipeline since they provided explicit and self-contained context, potentially aiding the model’s generalization and interpretability. Lastly, we evaluated whether switching the prompt language from Spanish to English afected performance. This comparison revealed negligible diferences (less than 0.1 points diference), leading us to retain Spanish for practical reasons, such as ease of human evaluation, reduced code-switching artifacts, and simplified prompt quality control.
Once we had identified the best-performing prompt for Gemma-3, fine-tuning Gemma-3 was infeasible within the competition’s tight timeline, we instead supervised-fine-tuned Gemini-1.5 with the winning prompt from the previous phase. The resulting model became the core of our system.
Using this fine-tuned Gemini-1.5, we conducted temperature sweeps and post-processing experiments. We found that = 0.80 produced the best results, whereas the optimal temperature for base Gemma-3 had been = 1.0. This upgrade lifted BLEU on the development set from 0.2978 (base Gemma-3 + best prompt) to 0.4201.
With the writer model finalized, we next generated candidate outputs and re-ranked them with a second Gemini-1.5 acting as a judge. Generating six candidates and selecting the best one raised BLEU on the test set by +4.7 points compared with single-shot decoding. Manual probing showed that a high-quality spoiler often appeared only on the third, fourth, or even sixth attempt; the first answer frequently missed key facts or omitted information. By letting the writer explore a small diversity of outputs and allowing the judge to pick the most concise and faithful candidate, we systematically increased the chance of capturing the correct answer gap.
Our final spoiler-generation model used a Gemini-1.5-flash-002 base, fine-tuned in supervised mode for 10 steps, with a learning-rate multiplier of 1 and an adapter size of 8, without intermediate checkpoints. A simplified code implementation with the winning model is available in a companion notebook 2.
This study introduced a simple approach to Spanish clickbait detection and a competitive architecture for spoiler generation, ranking fourth in classification and first in spoiling. Supplying the language model with the full, cleaned article consistently outperformed any information-retrieval strategy that restricted context, because crucial details often appear late in the body and a truncated window pushed the model toward spurious “No hay respuesta” replies. A small, manually balanced set of Spanish examples, each enriched with an explicit chain of thought, guided the model more efectively than retrieval-based prompts whose similarity was measured only at the lexical level; the benefit stems from showcasing diverse reasoning patterns rather than surface overlap. Finally, replacing the base Gemma-3 with a supervised fine-tuned Gemini-1.5 and re-ranking multiple candidate outputs with a second “judge” instance yielded a cumulative gain of more than 13 BLEU points. This shows the value of supervised fine-tuning.
Future research could explore multimodal LLMs: some training tweets contained spoilers available only through an embedded image (see Table 8), we removed them from the train and val dataset before performing the fine-tuning. Injecting an automatic caption of that image obtained with a multi-modal LLM into the textual context could unlock currently unsolved cases. Most prompt-engineering choices and decoding hyper-parameters (temperature, number of candidates, prompt language) were tuned on the intermediate Gemma-3 model and only quick-checked once the pipeline migrated to the finetuned Gemini-1.5 writer + judge. A systematic, model-specific search (e.g., grid search) directly on Gemini—could reveal combinations better aligned with the newer model’s behavior and yield additional gains in both automatic scores and human-rated fluency.
During the preparation of this work, the authors used two generative-AI tools—OpenAI GPT-4o and OpenAI o3— for copy-editing tasks (grammar and spelling). All AI-suggested edits were reviewed 2https://colab.research.google.com/drive/1qnfduoTNdUTE-Tnb8kCZfFGR-ujwWiOW?usp=sharing and, where necessary, modified by the authors, who accept full responsibility for every aspect of this publication. G. de Rosa, O. Saarikivi, A. Salim, S. Shah, X. Wang, R. Ward, Y. Wu, D. Yu, C. Zhang, Y. Zhang, Phi-4 technical report, 2024. URL: https://arxiv.org/abs/2412.08905. arXiv:2412.08905. [19] P. Fu, Y. Zhang, H. Wang, W. Qiu, J. Zhao, Revisiting the knowledge injection frameworks, in: H. Bouamor, J. Pino, K. Bali (Eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Singapore, 2023, pp. 10983–10997. URL: https://aclanthology.org/2023.emnlp-main.677/. doi:10.18653/v1/2023. emnlp-main.677.