01 / AI · Analysis
TeViD compares two approaches to Turkish video fact-checking
TeViD combines fact-checking articles, translated material and synthetic examples to compare evidence retrieval with a fine-tuned model.

Azra Açıl presented TeViD, a Turkish video fact-checking dataset, in a paper comparing two verification systems. One retrieves supporting evidence; the other uses a model adapted through further training. Both share the same open-source base model.
The paper describes existing fact-checking systems as focused largely on English text. Tools for Turkish remain severely limited, according to the author. Açıl presents TeViD as the first Turkish multimodal video fact-checking dataset, to the best of the author's knowledge.
The comparison examines different ways to assess claims associated with video. The retrieval system divides verification tasks among several roles. The fine-tuned system produces verdicts through a single model without external searches.
Articles, translations and synthetic examples
According to the paper, TeViD draws on articles published by Teyit.org. It describes Teyit.org as an organisation certified by the International Fact-Checking Network (IFCN). Translated Snopes articles supplement that material.
Açıl also added synthetic examples to the dataset. These were intended to reduce the imbalance between classification categories. The abstract does not disclose the number of examples in each category.
Retrieved evidence and fine-tuned verdicts
The paper's retrieval system, 3MFact-TR, uses the Qwen2.5-7B language model and the Qwen2-VL-7B vision-language model. It adds modules for checking semantic inconsistencies, managing who, what, when, where, why and how questions, and retrieving evidence in two languages. The retriever excludes fact-checking sites to prevent answers from leaking into the evidence.
For the fine-tuned system, Açıl used a single Qwen2.5-7B model. Training used Low-Rank Adaptation (LoRA), a method that adjusts a limited set of model parameters. The abstract identifies the verdict inputs as the claim and a textual description of the video.
The paper identifies accuracy and macro F1, which averages F1 scores across categories, as evaluation measures. The abstract does not disclose performance figures for either system. It also names F1 and recall for the true category as measures for comparing their verdicts.