A Spanish Synthetic Text Dataset for Evaluating SLM-Generated Text under RAG and NO RAG Configurations

This dataset comprises synthetic news descriptions in Spanish, generated from real newspaper headlines using a controlled Large Language Model (LLM) pipeline. The generation process follows two distinct configurations: Retrieval-Augmented Generation (RAG) and generation without external context (NO RAG). For each configuration, three temperature settings (1.0, 0.75, and 0.5) were applied to control the variability and determinism of the generated outputs. The source headlines were collected over a two-week period from two major Spanish newspapers, referred to as Newspaper A and Newspaper B. In the RAG configuration, each synthetic description is generated using contextual information retrieved from a knowledge base composed of past news descriptions from the same newspaper. These historical texts were embedded and stored in a vector database, enabling semantic retrieval of relevant context for each generation instance. The dataset contains 61,374 synthetic descriptions in total (18,112 from A and 43,262 from B). The two generation configurations (RAG and NO RAG) are equally represented. To provide a clear overview of the dataset, the key details are summarized in the table below: Headlines collection period 26 October 2025 – 11 November 2025 RAG knowledge base period 25 September 2025 – 25 October 2025 Language Spanish LLM models Ministral-3B, Qwen-2.5-0.5B and Llama-3.2-1B Temperature values 1, 0.75 and 0.5 The dataset is delivered as a collection of JSON files organized in a hierarchical folder structure for clarity and ease of use. The folder structure is as follows: dataset/ + 01. llm_news_NO_RAG/ - llm_news_NO_RAG_A.json - llm_news_NO_RAG_B.json + 02. llm_news_RAG/ - llm_news_RAG_A.json - llm_news_RAG_B.json The subfolders distinguish between the generation configurations: llm_news_NO_RAG contains descriptions generated without contextual retrieval. llm_news_RAG contains descriptions generated using retrieval augmented generation. Each JSON file corresponds to a specific newspaper (A or B) and includes all generated descriptions for that source. Each entry of the JSON file corresponds to a single generated description. Each JSON entry includes metadata fields that describe the experimental configuration, such as: RAG 0: Without retrieval 1: With retrieval timestamp_llm indicates when the description was generated. id_llm identifies the model configuration used during generation. 1: Ministral-3B 2: Qwen-2.5-0.5B 3: Llama-3.2-1B id_feature encodes the temperature (t) parameter used during generation. The mapping is as follows: 1 → t = 1.0 2 → t = 0.75 3 → t = 0.5