KomMKonLLM - Combinatorial test cases for consistency testing of LLMs

This dataset was created using the KomMKonLLM implementation available at https://github.com/KomMKonLLM/KomMKonLLM. Example dataset and results This folder contains an example set of questions and labelled correct answers, the synonyms utilized in the testing process (which forms the basis for the construction of input parameter models required to generate a covering array), the resulting covering array, and the LLM output for 6 models as well as its boolean interpretation. Models The following models are used in this example, all interfaced with using the Ollama meta-model: Starling-LM Llama 3.1 Llama 3.2 Llama 2 Uncensored DeepSeek-r1 Mistral Questions The file `public_questions.jsonl` contains 27 questions and the associated correct answer label. It was constructed using a human-guided ChatGPT session. Test runs Data regarding the executed test runs is available in `test_runs.csv`.The first column contains a unique ID that is also utilized in the file names for synonyms and covering arrays described below.Each row further contains the original sentence text, the date, the correct answer, the model name (which is always OLLAMA in these examples), the note (which contains the actual model) and the strength.Note that identical sentence texts are contained multiple times in this file, once per model. Synonyms Files named `synonyms-<id>.txt` contain the synonyms for the test sentence with the ID `<id>`.Each line is a JSON list with synonyms for a token (roughly equivalent to a word). A limit of 3 synonyms per token is applied internally and only some types of tokens (such as proper nouns and adjectives) are replaced with synonyms.The first entry in each of the lines is the token as it appears in the original sentence, with other entries provided by a NLP library. Covering Arrays The generated covering arrays (at strength 2 in this example) are available in files called `ca-<id>.csv`. They are based on the associated synonym list such that the n-th column in `ca-<id>.csv` refers to the n-th row in `synonyms-<id>.txt` and each row in the CA file forms one test case. Queries The resulting queries (i.e. mutated test sentences) are listed in `queries.csv`. This file contains the full prompt submitted to the LLM (note that the `sentence_id` in this file refers to the `id` in `test_runs.csv`; this can be used to identify the concrete LLM that was queried), the LLM's response, and the response interpreted as a boolean via our oracle.In each of these rows, the value of the `correct_answer_label` column should be compared to the `result` column to evaluate whether the LLM's response was correct. Results This section visualizes the results of our analysis, which was performed using the JupyterLab notebook contained in this project. Parsed LLM responses The chart in llm-responses-parsed.png shows the parsed LLM responses across all test runs. Responses that could not be parsed as boolean are listed as "undefined". In general, this means that the LLM did not return a JSON boolean. Some models are prone to printing significantly more complex data structures instead or try to add a textual explanation instead of following the prompt. Precision The chart precision.png visualizes the precision of each model in this analysis.The precision is defined as the ratio of true positives over the sum of true positives and true negatives. Recall The file recall.png shows the recall of the models contained in our evaluation.The recall is the ratio of true positives over the sum of true positives and false negatives. F1 score The image f1.png contains the F1 scores of all models under test, defined as 2*(precision*recall)/(precision + recall). Consistency The chart in consistency.png visualizes the consistency of parsed LLM responses within each test run.The consistency measures the ratio of responses to mutated queries (i.e. those where parts have been replaced with synonyms) that are identical to the response of the original query.