This paper presents the approach developed by the UCOM-UNAM-PLN team for Task 3 of the CLEF 2025 CheckThat! Lab, focused on verifying numerical and temporal claims in Spanish. We propose a two-step factchecking architecture leveraging large language models (LLMs) in both stages: (1) evidence retrieval through BM25-preselected documents, and (2) veracity classification using zero-shot and few-shot prompting strategies. We conduct a comprehensive evaluation of retrieval techniques-including statistical, embedding-based, and LLM-based models-on the CLEF 2024 development set, identifying GPT-4o as the top-performing retriever. Based on these findings, we deploy GPT-4o for both evidence extraction and classification. Our best-performing pipeline combines few-shot prompting with curated evidence, achieving a macro-averaged F1 score of 0.3595 on the oficial CLEF 2025 test set. Results highlight the efectiveness of hybrid LLM-based architectures in identifying false claims, while also revealing challenges in handling ambiguous or true statements. This study underscores the potential of combining semantic retrieval and task-specific prompting for robust fact verification in real-world contexts.
In their state-of-the-art review, [3] organize automated fact-checking methods according to four-stage
architecture proposed by [4]: (
In their review on automated fact-checking, [3] identify three main approaches for evidence retrieval: traditional lexical-based methods like BM25, dense embedding-based methods, and hybrid approaches that combine initial retrieval and re-ranking with more complex models. For information retrieval, [5] employs a transformer-based model inspired by extractive question-answering systems. This model, typically BERT-based, takes a claim and a reliable source (such as a Wikipedia article) and extracts a relevant text snippet for claim verification. The training is done by fine-tuning pre-trained BERT models (such as BERT and DistilBERT) using the FEVER dataset. In the experiments conducted, DistilBERT achieved an exact match score of 90.19% and an F1-score of 93.98%, slightly outperforming BERT, which obtained an exact match score of 89.89% and an F1-score of 93.93%.
In [5] the claim verification model is based on BERT. This model takes as input the original textual claim and the retrieved textual evidence. Sentence embeddings for the claim and the evidence are obtained through a pre-trained BERT embedding layer. These embeddings are then processed and passed to a fully connected layer. Finally, the output of this layer is used to classify the claim into one of three categories: true, false, or "not enough information" (NEI), determining whether the evidence is suficient to verify the claim. In the verification stage in [ 6], focuses primarily on the use of Large Language Models (LLMs). These LLMs, such as GPT-4 and Llama3-70B-Instruct, are guided by system prompts to analyze a claim and the provided textual evidence. Their goal is to predict a label (SUPPORTS, REFUTES, or NOT ENOUGH INFO) along with a confidence score, leveraging their deep understanding of natural language for complex reasoning tasks. In [7], the authors conducted a comprehensive analysis on the identification of critical statements using various strategies, including lexical representations, embeddings, and LLMs. Specifically, they evaluated the use of LLMs as direct classifiers by applying zero-shot and few-shot prompting strategies, without requiring additional training.
Large Language Models (LLMs) are emerging as a promising component in fact-checking systems. In this work, we explore the use of Large Language Models (LLMs) in both stages of the fact-checking architecture.
1 { 2 3 4 5
To address this challenge we only used the development dataset of the Checkthat 2025 Lab [2] which contains elements in JSON format with the following structure: " c l a i m ": " San L u i s no t i e n e e m p l e a d o s pú b l i c o s en e x c e s o ; hoy son 17 m i l ", " c r a w l e d _ d a t e ": " 2011 −09 −26 ", " c o u n t r y _ o f _ o r i g i n ": " a r g e n t i n a ", " doc ": " L a s c i f r a s d e l P r e s u p u e s t o p r o v i n c i a l son c o n s i s t e n t e s con e s a a f i r m a c i ón . No a s í l o s r e s u l t a d o s de l a E n c u e s t a Permanente de Hogares . O t r o s i n d i c a d o r e s m u e s t r a n que l a p r o v i n c i a t i e n e una a l t a t a s a de empleo en n e g r o y de p l a n e s s o c i a l e s . San L u i s e s [ l a p r o v i n c i a ] mejor a d m i n i s t r a d a de l a Repú b l i c a porque no t i e n e e m p l e a d o s pú b l i c o s en e x c e s o ; hoy son 17 m i l e
The test dataset with 1.806 claims has a diferent structure than the development dataset with 377
claims. This dataset contained only the claims without associated documents. For this reason, it was
necessary to perform an additional task to collect the corresponding evidence documents for each claim.
To obtain the evidence, we used the corpus provided by the challenge organizers. In this context, we
considered two alternatives: (
Table 1 presents the statistic of the development. It shows class distribution of the development dataset, documents length, and the average evidence length measured in characters.
From each entry of the development dataset, we extracted: the claim, the doc (the document containing the associated evidence) and the label.
Our proposed approach is based on the use of large language models (LLM), which were employed in
two diferent ways: (
Both approaches were configured under two diferent inference schemes: zero-shot, in which the model performs the task without having received prior examples during the interaction, and few-shot, where representative examples are provided within the prompt to guide its behavior and attempt to improve accuracy in classification or evidence extraction.
The following section provides a detailed description of both stages of the approach.
In this first approach, the LLM was instructed to act as an automatic claim classifier. Its goal was to analyze each claim along with its associated evidence and assign a veracity label based on the provided content. The instructions used are shown below: You are an a s s i s t a n t focused on f a c t v e r i f i c a t i o n involving numerical claims and temporal expressions . Your r o l e i s to c l a s s i f y the t r u t h f u l n e s s of each claim based on the evidence provided .
Numerical claims are defined as those req uiring v a l i d a t i o n of e x p l i c i t or i m p l i c i t q u a n t i t a t i v e or temporal d e t a i l s . You w i l l r e c e i v e d a t a i n JSON format . Each i n s t a n c e w i l l i n c l u d e : − a c l a i m ( t h e s t a t e m e n t t o v e r i f y ) , − and an e v i d e n c e ( t h e s u p p o r t i n g or c o n t r a d i c t i n g i n f o r m a t i o n ) , You must a n a l y z e t h e e v i d e n c e and a s s i g n one o f th e f o l l o w i n g l a b e l s t o th e c l a i m : − True − F a l s e − C o n f l i c t i n g Only respond with t h e l a b e l . No a d d i t i o n a l e x p l a n a t i o n i s r e q u i r e d .
In the case of the few-shot approach, the same base instructions as in the zero-shot scenario were used, but labeled examples were added to the prompt in order to guide the model in the task. To do this, the following instruction was included: You w i l l f i r s t r e c e i v e a few l a b e l e d examples . Use t h e s e examples as gu idanc e t o d e t e r m i n e t he c o r r e c t l a b e l f o r t he f o l l o w i n g u n l a b e l e d i n s t a n c e s . Examples :
Following this instruction, an example of a claim with its corresponding evidence and label was included for each of the possible classes (True, False, and Conflicting ). These examples were intended to illustrate both the expected format and the necessary criteria for the model to perform a coherent classification aligned with the task. It is worth noting that, in order to identify the examples to be used, a manual curation was carried out to obtain representative examples.
In a second approach, the LLM was instructed as an evidence retriever, with the goal of identifying relevant fragments within the document that could later be used in the classification of the claim. Specifically, the model was asked to extract three sentences from the doc field, assigning each one a label indicating its relation to the claim (True, False, or Conflicting ), depending on whether the sentence supports, contradicts, or shows ambiguity with respect to the claim. The instructions used for the zero-shot case are provided below:
In the case of the few-shot approach, the same base instructions as in the zero-shot scenario were used, but labeled examples were added to the prompt in order to guide the model in the task. To do this, the following instruction was included: Here a r e some examples . Use t h e s e examples as a guide t o g e t t h e c o r r e c t s e n t e n c e s .
Once the relevant sentences were obtained, the following logic was applied to determine the final label of the claim: • If at least one of the sentences was labeled as Conflicting , the claim was classified as
• In the absence of Conflicting
sentences, if at least one sentence was labeled as False, the claim was classified as
• If all the sentences were labeled as True, the claim was classified as
• In cases where the retriever failed to identify relevant fragments and therefore returned no sentences, the claim was classified as
False by default.
In this work, we adopted this heuristic under the assumption that the absence of relevant evidence suggests that the claim is false. If a claim were true or conflicting, it is likely that verifiable information associated with it would exist in the document corpus. Therefore, a complete lack of retrieved evidence could lead us to conclude that the claim is false in this context. It is possible that there are cases in which no evidence exists—either because the evidence retriever failed to find the relevant evidence or because such evidence simply does not exist. In either case, a default class must be chosen at the time of classification. We acknowledge that an alternative could be to classify these cases as “not enough information for a verdict,” however, this class is not used in the current system. It is clear that this heuristic may introduce bias toward the False class. Nevertheless, introducing a new class like NOT ENOUGH INFO could also introduce bias toward that class, since the classification stage strongly depends on the evidence retrieval stage. 4.2.1. Use of evidence provided by the organizers The organizers provided a file containing the claims along with the 100 most relevant pieces of evidence retrieved from the corpus using the BM25 algorithm. Each entry in the file had the following structure: 1 { 2 3 4 5 6 7 } "doc_id": [list of document identifiers], "scores": [list of scores associated with each selected evidence], "query_id": "claim identifier number", "claim": "claim", "docs": [list of corresponding evidence]
For the experiments, the five pieces of evidence with the highest scores were selected for each claim. These pieces of evidence were concatenated into a single text, which served as the input document (doc). This document, along with the corresponding claim, was sent to the assistant to assign the veracity label or to extract the most relevant evidence.
Before addressing Task 3 of the CheckThat! 2025 challenge — Fact-Checking of Numerical Claims — , we ran preliminary experiments using the dataset from the CheckThat! 2024 Rumor Verification task [8]. The goal of this task was to explore information retrieval techniques and select the most suitable one as the relevant evidence retriever.
We conducted a comprehensive evaluation of a wide range of retrieval techniques, grouped into three main categories: • Traditional statistical techniques [9, 10, 11, 12, 13, 14, 15, 16]:
BM25, BM25PLUS, BM25-OKAPI, BM25+PL2, BM25LARGE, PL2, TF-IDF, DPH, DPH+PL2, DLH, LEMUR, HIEMSTRA, DFIZ, DFIC, DFR-BM25. • Embedding-based models [17, 18, 19] :
OpenAI : text-embedding-3-large, text-embedding-3-small.
Gemini: text-embedding-005-FACT-VERIFICATION, text-embedding-005-SEMANTICSIMILARITY, text-embedding-005-RETRIEVAL-DOCUMENT.
SBERT : SBERT-all-MiniLM-L6-v3. • Large Language Models (LLMs) used as semantic retrievers [20, 21, 22, 23]: GPT-4o: gpt-4o-2024-08-06, gpt-4o-mini-2024-07-18.
Qwen: qwen2.5-72b.
LLAMA: LLAMA3.2LARGE.
Deepseek: deepseek-llm-67b.
These techniques were evaluated on the CheckThat! 2024 development set, which contains 32 rumors, using Recall@5 (R@5) [24] which measures the model’s ability to retrieve at least one relevant document among the top 5 results and Mean Average Precision (MAP) which measures the model’s ability to
MAP rank relevant documents higher across the entire result list as evaluation metrics. The key results are summarized in Table 2. Bellow we draw some conclusions from these results.
LLMs as Retrievers. Large Language Models (LLMs) demonstrated a remarkable improvement in semantic retrieval capabilities. Notably, GPT-4o-2024-08-06 achieved the best overall performance, with R@5=0.940 and MAP=0.918, significantly outperforming all other techniques. This performance indicates a deep semantic understanding of claims and an outstanding ability to locate relevant evidence, even when it is expressed implicitly or through paraphrasing. Other LLMs, such as Qwen 2.5-72B and GPT-4o-mini, also delivered strong results (MAP > 0.74), albeit with slightly lower recall. In contrast, models like DeepSeek-67B and LLAMA3.2 Large showed significantly lower performance, possibly due to limitations in their training or in how they represent truthfulness or evidential relations in retrieval tasks.
Embedding-based Models. Pretrained embedding models ofered an interesting balance between eficiency and quality. The OpenAI text-embedding-3-large model achieved notable results (R@5=0.800, MAP=0.731), approaching the performance of some LLMs while maintaining a lower computational cost. Within this category, Gemini-FACT_VERIFICATION remained competitive (MAP = 0.695), confirming its suitability for contextual verification tasks. SBERT also yielded acceptable results considering its lightweight nature.
Traditional Statistical Techniques. While often considered baselines, traditional statistical retrieval methods such as BM25, PL2, TF-IDF, and DFR variants demonstrated surprisingly competitive performance, particularly when evaluating MAP. Notably, DFIZ achieved a MAP of 0.680 and DPH reached 0.676, both comparable to modern embedding models such as GEMINI-fact-verification (MAP=0.695), GEMINI-semantic-similarity (MAP=0.685), and OpenAI-embedding-3-small (MAP=0.674). In fact, DFIZ and DPH outperformed other embedding-based retrievers like GEMINI-retrieval-document (MAP=0.670), showing that statistical methods can still ofer strong baseline performance under the right conditions. These results suggest that, despite lacking semantic understanding, term-frequency-based approaches remain viable for certain fact-checking tasks—particularly when the evidence is lexically similar to the claim. Their lower computational cost and strong MAP scores make them attractive for large-scale or resource-constrained settings.
To address Task 3 of the CheckThat! 2025 Lab, which consists in verifying numerical and temporal claims in Spanish, we implemented and evaluated two distinct pipelines using the gpt-4o-2024-08-06 model. This model was selected based on its outstanding retrieval performance in the preliminary experiments (MAP = 0.918, R@5 = 0.940), as shown in Section 5.
We explored two main strategies: • Direct classifier: a single-stage prompt that directly predicts the veracity label (True, False, or
Conflicting) given the claim and evidence. • Evidence collection + classification: a two-stage pipeline where evidence is first retrieved using a semantic retriever (top-5 from BM25 preselection), and then the claim is classified based on this curated evidence.
Each strategy was evaluated using two prompting approaches: • Zero-shot: the LLM receives only the task instruction.
• Few-shot: the LLM is provided with a small number of labeled examples before prediction.
The experiments were carried out using the OpenAI assistant configuration, with a top p value set to 1 and a temperature of 0.01. This setup minimizes the randomness of the generated responses, making the model highly deterministic by prioritizing the most probable tokens at each step, while still considering the full probability distribution due to the top-p value of 1.
We report macro-averaged F1 [25] score on the oficial Spanish development set. The results are summarized in Table 3. This metric was selected as it aligns with the oficial evaluation criteria used by the competition organizers, ensuring consistency and comparability with the leaderboard results.
The few-shot evidence+classifier pipeline yielded the best result (F1 = 0.672), indicating that combining relevant evidence selection with prompting examples significantly improves performance. In contrast, the zero-shot direct classification approach underperformed, achieving only F1 = 0.349, highlighting the importance of both evidence quality and task-specific guidance.
In the Checkthat 2025 Lab, we participated in Task 3 (Fact-Checking Numerical Claims) for the Spanish dataset. According to the competition rules, teams were allowed to submit only one final run for evaluation on the oficial test set.
Based on our validation results (see Table 3), we selected the Evidence + classifier strategy with few-shot prompting using the gpt-4o-2024-08-06 model, as it achieved the best macro F1 (0.672) on the development set.
The final submission was evaluated by the organizers using macro-averaged F1, as well as classspecific F1 scores.
Our approach achieved the third-best macro F1 score (0.3595) in the oficial CheckThat! 2025 Task 3 evaluation for Spanish (see Table 4). While not leading in overall ranking, our model demonstrated particularly strong performance in identifying False claims, with an F1 score of 0.7443, indicating a reliable ability to detect and reject incorrect numerical or temporal assertions. However, the F1 scores for the True (0.1853) and Conflicting (0.1490) labels were significantly lower on the test dataset. This disparity in class-wise performance suggests a particular challenge in distinguishing and handling claims that are not clearly "False", an aspect we address in more detail in the limitations and future perspectives of this work.
Despite the promising results obtained by the proposed pipeline, several limitations must be acknowledged. The computational cost associated with using large language models like gpt-4o could be a practical limitation for large-scale deployments or resource-constrained environments, despite the observed performance gains. This limitation highlights the need to explore smaller models in future work.
Furthermore, the efectiveness of the LLM-based components, particularly in the few-shot configuration, relies heavily on the quality and representativeness of the examples provided; poorly chosen examples can lead to suboptimal classifications or evidence extraction. The classification is highly dependent on the evidence retrieved; in cases where relevant evidence is not retrieved by BM25, the system may fail to properly assess the claim’s veracity.
A crucial limitation that explains the lower F1 scores for the "True" and "Conflicting" labels lies in the classification heuristic applied. When the evidence retriever identifies relevant sentences, the current logic assigns specific priority to the verdicts of the extracted sentences. If at least one sentence is "Conflicting", the claim is classified as "Conflicting". In the absence of "Conflicting" sentences, if at least one sentence is "False", the claim is classified as "False". Only if all sentences are "True" is the claim classified as "True". This precedence implies that if sentences with both "True" and "False" verdicts are retrieved simultaneously (i.e., contradictions exist among the relevant sentences) and no sentence is explicitly labeled as "Conflicting" by the LLM, the system tends to classify the claim as "False" due to the priority given to the "False" verdict. This may result in the failure of identification of "Conflicting" or even "True" claims when the retrieved evidence is not uniformly positive.
Likewise, in cases where the retriever failed to identify relevant fragments and therefore returned no sentences, the claim was classified as False by default. This heuristic, adopted under the assumption that the absence of verifiable evidence suggests the claim is false in this context, further biases the system toward the "False" classification, contributing to the poor performance in the "True" and "Conflicting" categories and potentially reducing the overall robustness of the classification.
This paper presented an efective two-step architecture for numerical fact-checking, leveraging the power of large language models within a hybrid pipeline. Our approach, combining semantic evidence retrieval with a few-shot classification strategy using the gpt-4o-2024-08-06 model, achieved a competitive macro F1 score (0.3595) and ranked third in the CheckThat! 2025 Task 3 oficial test set for the Spanish dataset [1]. The results underscore the significant impact of evidence quality and task-specific guidance through few-shot prompting on fact verification performance.
For future work, several promising avenues remain open for exploration. One potential direction is to investigate alternative evidence retrieval strategies, including the full implementation and evaluation of embedding-based retrieval methods using dedicated performance metrics. Such approaches could enrich the initial evidence pool and potentially lead to improved overall system performance.
In addition, to more efectively address the identified limitations in the classification of "True" and "Conflicting" claims, there is a need to improve the classification heuristic. Specifically, we will explore a more sophisticated logic to handle situations where the retrieved sentences present contradictions (e.g., a mix of "True" and "False" verdicts without an explicitly "Conflicting" sentence). The goal is to ensure that such cases are more accurately classified as "Conflicting", better reflecting the inherent ambiguity of contradictory evidence.
Finally, given the computational cost and scalability limitations of large models like GPT-4o, exploring the potential of smaller language models is identified as a priority for future work. This line of research will include both a comparative evaluation and the application of fine-tuning techniques specifically tailored to fact-checking tasks.
Fine-tuning lighter models will significantly reduce deployment costs and also open up the possibility of training models to capture complex semantic nuances. For example, in situations where all retrieved evidence appears to be "True" or "False", but the claim contains a subtlety that requires a "Conflicting" classification, a properly fine-tuned model could learn to recognize these patterns beyond traditional heuristic rules. By incorporating such subtleties into supervised training, we expect to enhance the system’s ability to produce more accurate and robust judgments.
These future work directions are considered fundamental steps to increase both the technical viability and the practical applicability of the proposed architecture, especially in contexts where eficiency is required without sacrificing accuracy.
This work was partially supported by UNAM PAPIIT project IN104424 and by the Mexican Government through SECIHTI Proyect FC-2023-G-64.
During the preparation of this work, the authors used chatGPT in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. on Research and Development in Information Retrieval, SIGIR ’24, Association for Computing Machinery, New York, NY, USA, 2024, p. 650–660. doi:10.1145/3626772.3657874. [3] Z. Guo, M. Schlichtkrull, A. Vlachos, A survey on automated fact-checking, Transactions of the
Association for Computational Linguistics 10 (2022) 178–206. doi:10.1162/tacl_a_00454. [4] A. Vlachos, S. Riedel, Fact checking: Task definition and dataset construction, in: C. DanescuNiculescu-Mizil, J. Eisenstein, K. McKeown, N. A. Smith (Eds.), Proceedings of the ACL 2014 Workshop on Language Technologies and Computational Social Science, Association for Computational Linguistics, Baltimore, MD, USA, 2014, pp. 18–22. URL: https://aclanthology.org/W14-2508/. doi:10.3115/v1/W14-2508. [5] A. Adesokan, S. Elbassuoni, Factify: An automated fact-checker for web information, in: 2024 IEEE International Conference on Big Data (BigData), 2024, pp. 1546–1551. doi:10.1109/ BigData62323.2024.10825147. [6] L. Kolb, A. Hanbury, Authev-lkolb at checkthat! 2024: a two-stage approach to evidence-based social media claim verification, Faggioli et al.[22] (2024). [7] G. E. Pianciola Bartol, A. Tommasel, Towards automated fact-checking: An exploratory study on identifying check-worthy phrases for verification, in: 2024 L Latin American Computer Conference (CLEI), 2024, pp. 1–10. doi:10.1109/CLEI64178.2024.10700241. [8] A. Barrón-Cedeño, F. Alam, T. Chakraborty, T. Elsayed, P. Nakov, P. Przybyła, J. M. Struß, F. Haouari, M. Hasanain, F. Ruggeri, X. Song, R. Suwaileh, The clef-2024 checkthat! lab: Check-worthiness, subjectivity, persuasion, roles, authorities, and adversarial robustness, in: N. Goharian, N. Tonellotto, Y. He, A. Lipani, G. McDonald, C. Macdonald, I. Ounis (Eds.), Advances in Information Retrieval, Springer Nature Switzerland, Cham, 2024, pp. 449–458. [9] A. Trotman, A. Puurula, B. Burgess, Improvements to bm25 and language models examined, in: Proceedings of the 19th Australasian Document Computing Symposium, ADCS ’14, Association for Computing Machinery, New York, NY, USA, 2014, p. 58–65. doi:10.1145/2682862.2682863. [10] Kocabaş, B. T. Dinçer, B. Karaoğlan, A nonparametric term weighting method for information retrieval based on measuring the divergence from independence, Information Retrieval 17 (2013) 153–176. doi:10.1007/s10791-013-9225-4. [11] G. Amati, E. Ambrosi, M. Bianchi, C. Gaibisso, G. Gambosi, Fub, iasi-cnr and university of tor vergata at trec 2007 blog track, 2007. [12] G. Amati, Frequentist and bayesian approach to information retrieval, 1970, pp. 13–24. doi:10.
1007/11735106_3. [13] B. He, I. Ounis, Term frequency normalisation tuning for bm25 and dfr models, in: D. E. Losada, J. M. Fernández-Luna (Eds.), Advances in Information Retrieval, Springer Berlin Heidelberg, Berlin, Heidelberg, 2005, pp. 200–214. [14] S. Robertson, H. Zaragoza, The probabilistic relevance framework: Bm25 and beyond, Foundations and Trends® in Information Retrieval 3 (2009) 333–389. doi:10.1561/1500000019. [15] J. Perea-Ortega, M. García-Cumbreras, M. García-Vega, L. López, Comparing several textual information retrieval systems for the geographical information retrieval task, volume 5039, 2008, pp. 142–147. doi:10.1007/978-3-540-69858-6_15. [16] D. Hiemstra, Using language models for information retrieval, 2001. [17] OpenAI, text-embedding-3: Openai embedding models, 2024. URL: https://platform.openai.com/ docs/guides/embeddings, accessed: 2025-05-28. [18] G. Research, Generalizable embeddings from gemini, arXiv preprint arXiv:2503.07891 (2025). URL: https://arxiv.org/abs/2503.07891. [19] N. Reimers, I. Gurevych, Sentence-BERT: Sentence embeddings using Siamese BERT-networks, in: K. Inui, J. Jiang, V. Ng, X. Wan (Eds.), Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Association for Computational Linguistics, Hong Kong, China, 2019, pp. 3982–3992. doi:10.18653/v1/D19-1410. [20] OpenAI, GPT-4 technical report, CoRR abs/2303.08774 (2023). doi:10.48550/ARXIV.2303.
08774. arXiv:2303.08774. [21] Qwen, et al., Qwen2.5 technical report, 2025. URL: https://arxiv.org/abs/2412.15115.
arXiv:2412.15115. [22] A. G. et al., The llama 3 herd of models, 2024. URL: https://arxiv.org/abs/2407.21783.
arXiv:2407.21783. [23] D.-A. et al., Deepseek llm: Scaling open-source language models with longtermism, 2024. URL: https://arxiv.org/abs/2401.02954. arXiv:2401.02954. [24] A. Yates, R. Nogueira, J. Lin, Pretrained transformers for text ranking: Bert and beyond, SIGIR ’21, Association for Computing Machinery, New York, NY, USA, 2021, p. 2666–2668. doi:10.1145/ 3404835.3462812. [25] F. Sebastiani, Machine learning in automated text categorization, ACM Computing Surveys 34 (2002) 1–47. doi:10.1145/505282.505283.