AI-Koditex

AI-Koditex is a generated, annotated, multi-genre corpus of Czech texts produced by large language models (LLMs).

Modeled on the original Koditex Corpus, AI-Koditex was created to mirror its structure, genre balance, and linguistic diversity, enabling direct comparisons between human and machine-generated Czech texts. The corpus includes texts generated by 13 frontier LLMs from major developers (OpenAI, Anthropic, Meta, Alphabet, DeepSeek), each prompted using material from the original Koditex corpus to preserve consistency in genre and topic distribution. Like its human-authored counterpart, the AI Koditex corpus is a mixed corpus encompassing a broad genre range. All texts are tokenized, lemmatized, and morphologically and syntactically annotated following the Universal Dependencies standard, and are released in both plain text and CoNLL-U formats. AI-Koditex is a large-scale Czech LLM-generated resource, developed to support cross-linguistic and cross-model research in AI-generated language.

The original reference Koditex Corpus was originally available in vertical format from the Czech National Corpus infrastructure. The pre-processing pipeline involved several steps to prepare prompts suitable for generation. Clean texts and metadata were extracted from the verticals, and structural tags were standardized. Each text sample was divided into two portions to enable prompt-based generation:

• Prompt portion: The first 500 words (including punctuation) served as generation prompts

• Reference portion: The remaining text (approximately 1,500 words) provided human-authored comparison material

This segmentation strategy ensured that models received sufficient context for generation while maintaining substantial reference text for comparative analysis. Also, the context of 500 words left sufficient space in the context window even for older models (davinci-002 has maximum context of 2049 tokens, while 500 English words takes about 670 tokens).

Importantly, unlike the Koditex corpus, AI-Koditex contains written texts only, one sample per source text to avoid over-representation. The final dataset contains 676 text samples.

For each model, we generated two versions of each corpus: one using temperature 0 (deterministic generation) and one using temperature 1 (stochastic generation). However, we encountered mode collapse with the oldest model (davinci-002) at zero temperature, resulting in constant repetition of identical sentences. Additionally, this early model failed to produce coherent Czech text.

For base models operating in completion mode (davinci-002, GPT-3.5-turbo, Meta-Llama-3.1-405B), we used only the first portion of each source text as input, allowing the models to function as traditional language models for text prediction.

For instruction-tuned models, we employed minimal system prompts requesting long continuation of given text. Without such prompts, models' default helpful assistant persona emerged who typically attempted to analyze, summarize, or answer questions within the source text rather than continuing it. Language-specific challenges emerged during Czech texts generation. Some models refused to cooperate when given Czech system prompts, necessitating English system prompt as quoted below. Despite explicit instructions to generate Czech text, several models sometimes produced English or mixed-language outputs.

We used the following system prompt: Please continue the Czech text in the same language, manner and style, ensuring it contains at least five thousand words. The text does not need to be factually correct, but please make sure it fits stylistically.

To ensure reproducibility, we used random seed 42 for all OpenAI API calls. Unfortunately, other providers do not offer comparable deterministic generation options. Llama generations used 16-bit floating-point quantization (the highest available quality).

API responses were preserved in their entirety, including token probabilities and alternative tokens when available, to enable future analysis of generation uncertainty and model confidence.

Texts that were too short were removed. For instruction-tuned models, the original phrases such as „I'd be happy to continue“ were also removed.

We used Universal Dependencies for annotation, as UDPipe represents the state of the art for multi-level linguistic processing (including tokenization, lemmatization, syntax, and morphology). The resulting annotation is in the CoNLL-U format, which is a widely adopted standard compatible with most modern NLP tools.