aimgo/Simulacrum

Si machina nullo moderatore aberrasset...

About

This dataset features human authored text, mined from Acervus, annotated by multiple LLMs with 5 different span relations and 4 different categorical features. The organic examples are paired with machine authored modifications.

It also contains a number of hallucinated or adversarial Latin-like examples which pass lexical filters but are semantically or grammatically incoherent. Additionally, there are some 10,000 samples created by a variety of Lorem Ipsum generators.

These include highly adversarial samples which contain imagined Latin forms, no discernible syntax, or are completely incoherent.

Atrius Tubulus consul centum senatus consultum triumphi non sregio pugnāvis
 ob sidera accipiet aga specta dimidia parte multa vis sed

Pseudo Latin can be observed across multiple different registers, further complicating detection:

conglutinatus est in hodie restandum quod vobis Et terra venter noster Clamabat que
ad dixerunt Monstra calumniatus nos Dominus

Relations

• NLI Logical Entailment (premise-hypothesis)
• Question Answering (question-answer)
• Quote Attribution (quote-attribution)
• Dialogue e.g. chat (utterance-reply)
• Prompt-Response, e.g. temporal succession, narrative sequence, or thematic association (prompt-response)

Categorical

• Text Genre (ecclesiastical, historiography, law, treatise, poetry, rhetoric, general, satire)
• Named Entities
• UPOS and UFEAT Annotation
• Adversarial Highly-Latin-Like Text (Pseudo-Latin, Lorem Ipsum, & Hallucinated Latin that evaded aggressive lexical filtering)

Machine Modification

• Human Original vs. Machine Authored Back Translation (human-machine)
• Original vs. Machine Authored Translation (lat-eng)

Total rows: 154426 (train 139116, test 15310)

models:

Model	Count	Train	Test
gpt-6.1-sol	75668	68195	7473
claude-opus-5-5	66762	60139	6623
LoremGenerators	11996	10782	1214

categories:

Category	Count	Train	Test
treatise	37229	33532	3697
ecclesiastical	31834	28574	3260
historiography	22771	20497	2274
pseudo	21508	19359	2149
entity-recognition	13805	12450	1355
pos-morph-tagging	7594	6868	726
poetry	17974	16187	1787
general	8296	7547	749
law	7279	6525	754
rhetoric	6883	6304	579
satire	652	591	61

relations:

Relation	Count	Train	Test
machine-authorship	43289	38982	4307
machine-translation	38004	34229	3775
quote-attribution	10410	9380	1030
prompt-response	8817	7954	863
premise-hypothesis	7544	6783	761
question-answer	2205	1991	214
dialogue	1303	1167	136

entity mentions:

Entity type	Count	Train	Test
PERS	17976	16211	1765
LOC	5895	5292	603
GRP	4812	4337	475

Tagged tokens: 65161 (train 58856, test 6305)

UPOS tags:

UPOS	Count	Train	Test
NOUN	16065	14499	1566
VERB	13759	12434	1325
ADJ	5933	5383	550
PRON	5130	4636	494
ADV	4476	4082	394
ADP	4028	3626	402
CCONJ	3861	3476	385
AUX	2901	2641	260
PROPN	2782	2501	281
DET	2605	2319	286
SCONJ	1660	1483	177
PART	1545	1397	148
NUM	362	329	33
INTJ	52	48	4
X	2	2	0

Citation

If you use this, please cite:

@misc{mccarthy2026Simulacrum,
  author       = {McCarthy, Aidan},
  title        = {{Simulacrum}: Machine Authored and Annotated Latin},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/aimgo/Simulacrum}},
  note         = {Dataset}
}