This is the code repository for the ACL2024 paper "Blinded by Generated Contexts: How Language Models Merge Generated and Retrieved Contexts When Knowledge Conflicts?"

• Environment: pip install -r requirements.txt Also need to clone llama-2 official inference code git clone -b llama_v2 https://github.com/meta-llama/llama.git into the project path.
• Download prepared context and intermediate results to the project directory. link
• Build the CC dataset and calculate DiffGR. python construct_cc_and_analyse.py

The resulting DiffGR matches Table 5, depicting the bias after excluding the effects of parametric knowledge. By adjusting the parameters in construct_cc_and_analyse.py, you can obtain datasets with different combinations of readers and generators (7b-chat, 13b-chat, gpt-4-0613 or gpt-3.5-turbo-0613).

1. First, we need to prepare the retrieved context and generated context and ensure their lengths match, following scripts/prepare_contexts.md.

2. Next, LLM answers questions based on different contexts, following scripts/answer_with_different_contexts.md.

   • Without context: To determine LLM's parametric knowledge.
   • Using only one type of context (generated/retrieved): To determine what each type of context provides.
   • Using both generated and retrieved context simultaneously: To analyze which context LLM relies on.

3. Construct the CC dataset and analyze LLM preference: python construct_cc_and_analyse.py