Add model-based results for MedNLI, RadNLI for GPT-3.5 and GPT-4

#8

What are you reporting:

  • Evaluation dataset(s) found in a pre-training corpus. (e.g. COPA found in ThePile)
  • Evaluation dataset(s) found in a pre-trained model. (e.g. FLAN T5 has been trained on ANLI)

Evaluation dataset(s): Name(s) of the evaluation dataset(s). If available in the HuggingFace Hub please write the path (e.g. uonlp/CulturaX), otherwise provide a link to a paper, GitHub or dataset-card.

Contaminated model(s):

  • This PR reports negative results for GPT-3.5 and GPT-4.

Contaminated corpora: None

Contaminated split(s): 0% over train/dev/test (MedNLI) and dev/test (RadNLI).

Briefly describe your method to detect data contamination

  • Data-based approach
  • Model-based approach

Description of your method, 3-4 sentences. Evidence of data contamination (Read below):

  • (Method is same as PR 3)
  • The only difference between this implementation and the original paper's (Golchin and Surdeanu 2024) is that here multiple runs (3 runs) were performed on each available split; this was to make sure that results hold across different (identically-sized) random data partitions. In addition the models were accessed through Azure OpenAI (opt out of human review + HIPAA-compliant), following MIMIC's DUA. For reference, a sanitized version of the results that keeps the data index, label, outputs, and contamination evaluation results without original input sentences can be found here.
  • While there are potential positives identified by the ROUGE-based contamination detection method, the best performing (GPT-4 ICL) detector did not consider these instances to be true contaminations, therefore this PR reports negative results (0% contamination for all splits on both datasets based on the examined method).

Citation

Is there a paper that reports the data contamination or describes the method used to detect data contamination?

URL: https://openreview.net/forum?id=2Rwq6c3tvr
Citation:



@article

	{DBLP:journals/corr/abs-2308-08493,
author       = {Shahriar Golchin and
                Mihai Surdeanu},
title        = {Time Travel in LLMs: Tracing Data Contamination in Large Language
                Models},
journal      = {CoRR},
volume       = {abs/2308.08493},
year         = {2023},
url          = {https://doi.org/10.48550/arXiv.2308.08493},
doi          = {10.48550/ARXIV.2308.08493},
eprinttype    = {arXiv},
eprint       = {2308.08493},
timestamp    = {Thu, 24 Aug 2023 12:30:27 +0200},
biburl       = {https://dblp.org/rec/journals/corr/abs-2308-08493.bib},
bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Important! If you wish to be listed as an author in the final report, please complete this information for all the authors of this Pull Request.

  • Full name: Jenny Chim
  • Institution: Queen Mary University of London
  • Email: c.chim@qmul.ac.uk
Workshop on Data Contamination org

Thank you @j-chim !!! Merged :D

Iker changed pull request status to merged

Sign up or log in to comment