README: This README contains a list of patches (Look at: random_scripts) made to get various experiments to work

CodeSage Model Patch for sentence_bert.py

This document outlines a required manual patch to run the Codesage-base-v2 model within the CodeRagBench framework.

The Problem

The CodeRagBench benchmark fails to load the Codesage-base-v2 model using the default beir library. The model requires the trust_remote_code=True flag to be set during initialization, which the library does not do by default.

---

The Patch

To resolve this, a direct modification was made to the conda environment's installed library file.

• File Modified:

  miniconda3/envs/crag/lib/python3.10/site-packages/beir/retrieval/models/sentence_bert.py
  

• Change: On or around line 30, inside the __init__ method, the AutoModel.from_pretrained() call was updated to include trust_remote_code=True.

  # ... existing code ...
  class SentenceBERT:
      def __init__(self, model_path: Union[str, Tuple] = None, sep: str = " ", **kwargs):
          # ... existing code ...
          self.model = AutoModel.from_pretrained(model_path, trust_remote_code=True) # <--- CHANGE APPLIED HERE
          # ... existing code ...

---

⚠️ Important

This is a manual override applied directly to a file within an installed package.

If the beir-py package is ever reinstalled or the crag conda environment is recreated from scratch, this change will be lost. The patch will need to be manually reapplied to get the Codesage-base-v2 model working again.