Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals

Authors: Yair Ori Gat, Nitay Calderon, Amir Feder, Alexander Chapanin, Amit Sharma, Roi Reichart

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results demonstrate the excellent performance of CF generation models as model-agnostic explainers.
Researcher Affiliation Collaboration T Faculty of Data and Decision Sciences, Technion, IIT CColumbia University Data Science Institute, MMicrosoft Research India
Pseudocode No The paper describes methods and objectives using prose and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code: https://github.com/Yair Gat/causal-explanations
Open Datasets Yes We employ CEBa B (Abraham et al., 2022), the only established benchmark for evaluating causal explanation methods.
Dataset Splits Yes CEBa B contains two train sets, exclusive (N = 1463) and inclusive (which we do not use), development (N = 1672), and test (N = 1688) sets.
Hardware Specification No The paper mentions 'GPUs' in a general context but does not provide specific hardware details like exact GPU or CPU models used for experiments.
Software Dependencies No The paper mentions specific language models and APIs used (e.g., MPNet, Chat GPT, GPT-4) but does not provide version numbers for ancillary software libraries or frameworks required for replication.
Experiment Setup Yes We use τ = 0.1, train the causal representation models for 12 epochs... We use a learning rate of 5e-6 for training the causal representation models. We use a learning rate 1e-5 and a batch size 16 for the fine-tuned baselines, explained models and concept predictors.