Detect AI-generated lines of Dante's "Divina Commedia"

Dante Intrusion Detection Dataset A benchmark for intrusion detection in text streams: detecting a small fraction of LLM-generated verses injected into authentic lines from Dante's Divine Comedy. Introduced in Sparse, self-organizing ensembles of local kernels detect rare statistical anomalies. Contents Text data: lines-divina-commedia.txt: authentic lines from Dante's Divine Comedy (Italian), one line per row. generate_lines_divina_commedia_meta-llama-3-70b-instruct_topk0.1_temp1.6.txt: LLM-generated Dante-style lines, one line per row. Embedded data (384-dimensional embeddings of each line, computed with multilingual Sentence-BERT paraphrase-multilingual-MiniLM-L12-v2, sentence-transformers): lines-divina-commedia.npy: embedding of the authentic lines from Dante's Divine Comedy (Italian), one line per row. generate_lines_divina_commedia_meta-llama-3-70b-instruct_topk0.1_temp1.6.npy: embedding of the LLM-generated Dante-style lines, one line per row. Row i of each embedding file corresponds to line i of the matching txt file Data generation Synthetic lines were produced by repeatedly querying Meta-Llama-3-70B-Instruct with the prompt: "Write five stanzas randomly extracted from the full Dante's Divina Commedia (italian). Each stanza has equal probability of being chosen." Sampling parameters: temperature=1.6, top_p=0.1, presence_penalty=1.15 (We use an intermediate temperature combined with low top-p to confine the added randomness to a small set of plausible words, while the presence penalty we chose reduces repetition across stanzas. Since generation still exhibits recurrent patterns, in our applications the output was filtered to unique lines only. This makes detection harder: it cannot rely on exact duplicates, which are a clear mark of LLM generation. Usage Build a reference sample from real lines and a test sample by mixing real lines with a chosen fraction of generated lines (2–20% in the original study, with samples of 1000 lines). Apply a two-sample test or anomaly-detection method to the embeddings to detect and characterize the intrusion. The txt files allow direct inspection of flagged lines or re-embedding with other models. Citation If you use this dataset, please cite: Sparse, self-organizing ensembles of local kernels detect rare statistical anomalies.