InstructSrc: An Instruction-Following Dataset for Source Attribution of LLM-Generated Texts
Description
InstructSrc is an instruction-following benchmark dataset for fine-grained source attribution of large language model (LLM)-generated texts. It contains texts generated by six instruction-tuned language models—Qwen2.5-7B, Phi-4-mini, Mistral-7B, Llama-3.2-3B, Falcon3-3B, and Falcon-7B—under a unified task collection and generation protocol. The dataset is designed for closed-set source-model attribution and supports the evaluation of attribution methods under both standard and shifted generation settings. Fixed training, validation, and test splits are provided to ensure reproducible and fair comparisons. The dataset can support research on LLM-generated text provenance, source-model identification, content auditing, and robust machine-generated text attribution. This dataset accompanies the manuscript “ProDIF: Proxy-Induced Dynamic Fingerprints for Black-Box Source Attribution of LLM-Generated Texts.”
Files
Steps to reproduce
The dataset contains six JSON files corresponding to texts generated by Qwen2.5-7B, Phi-4-mini, Mistral-7B, Llama-3.2-3B, Falcon3-3B, and Falcon-7B. The texts were generated under a unified task collection and generation protocol. Each JSON file contains the generated texts and the associated information required for source-model attribution experiments. The dataset can be loaded using a standard JSON parser and used for training and evaluating closed-set LLM source-attribution models.