This is the repository for Mitigating Bias in RAG: Controlling the Embedder! This repository includes the GenderBias-QA and PoliticBias-QA datasets, as well as the training and inference code to reproduce our results. The goal of this repo is to bias the embedder for these datasets to achieve an unbiased RAG system.

• Getting Started
• Sorting Documents
• Fine-tuning an Embedder
• Evaluation and Inferencing
• Evaluating Projections

This repository includes:

• Datasets: GenderBias-QA and PoliticBias-QA
• Training Script: Fine-tune an embedder with PEFT or WiSE-FT to adjust its bias
• Evaluation Script: Evaluate the performance of the biased embedder within a full RAG pipeline
• Projection Script: Assess projections in the embedding space for biasing retrieved documents

First, install the Python packages by running:

conda env create --file environment.yml

Second, enter your API keys and cache directory in env_config.sh and run:

source env_config.sh

To prepare for contrastive fine-tuning, you should first sort positive and negative documents from a corpus. To do this, run:

bash run-bounds.sh

This command runs src/rag-get-bound.py which will save the indices of the positive and negative documents in dataset/filtered_bias_index. The saved indices will later be used during fine-tuning.

Here are some of the main arguments to be passed into rag-get-bound.py:

• -query_d: Chooses the bias dataset (either PoliticBias-QA or GenderBias-QA).
• -corpus_d: Specifies the corpus in which the documents get sorted from. Possible choices include:
  • mteb/msmarco
  • mteb/dbpedia
  • mteb/fever
  • mteb/nq
  • mteb/hotpotqa
  • webis/args_me
  • webis-argument-framing
  • webis/conclugen
  • Pol_NLI

After you have sorted your documents for training, you can fine-tune an embedder to change its bias by running:

bash run-train.sh

This command runs embedder/train.py to train the embedder and evaluate both its accuracy and bias. The evaluation results are saved to a JSON file. By default, the trained model is not saved unless you provide the --save_dir argument.

Here are some of the main arguments to be passed into train.py:

• --bias_dataset: GenderBias-QA or PoliticBias-QA.
• --lr: Learning rates to use (e.g., [3e-5, 1e-5]).
• --epochs: Number of epochs (e.g., [5, 10, 15]). Setting this to 0 will evaluate the model without training.
• --freeze_layers: Number of layers to freeze (e.g., [0, 1, 2, 3, 4]).
• --merge_weight: Weights for merging (e.g., [0.1, 0.3, 0.5, 0.7, 0.9]).
• --train_corpus: Specify training corpus parts by using a hyphen to merge names (e.g., nq-msmarco-conclugen). Allowed corpus names include: msmarco, fever, dbpedia, nq, hotpotqa, conclugen, argument, args_me.
• --save_dir: Optionally provide a directory to save the trained model; if omitted, only the accuracy and bias evaluations will be saved to a JSON file.

The fine-tuning code is adapted from SGPT.

To automatically print the command for fine-tuning multiple embedders with varying configurations (e.g., learning rate, epochs, numbers of frozen layers), modify experiments/make_commands.py and run:

cd experiments
python make_commands.py

This will create a text file with each line being a command to run one fine-tuning job.

After an embedder has been trained, you may test the performance of the embedder within a full rag pipeline. This will evaluate the bias of the embedder and RAG pipeline. To evaluate, run:

bash run-inference.sh

This command runs src/rag-inference.py which will evaluate the bias of a full RAG pipeline. You will need to provide the LLM, corpus, and embedder for evaluation.

Here are some of the main arguments to be passed into rag-inference.py:

• -query_d: Specifies the queries to retrieve documents for. Can be chosen between GenderBias-QA or PoliticBias-QA.
• -corpus_d: Specifies the corpus for inference. Can be chosen between mteb/nq, mteb/hotpotqa, or Pol_NLI.
• -llm: Specifies the Huggingface LLM to inference on
• -rm: Specifies the embedder to retrieve documents with. Select a base embedder from Huggingface or a directory to a trained embedder.

You can also bias embedders by projecting them in the embedding space instead of fine-tuning. To do so, run:

bash run-projection.sh

This command runs src/rag-projection.py which will evaluate the bias of the full RAG pipeline using projections in a base embedder (gte-base by default).

Here are some of the main arguments to be passed into rag-projection.py:

• -query_d: Specifies the queries to retrieve documents for. Can be chosen between GenderBias-QA or PoliticBias-QA.
• -corpus_d: Specifies the corpus for inference. Can be chosen between mteb/nq, mteb/hotpotqa, or Pol_NLI.
• -llm: Specifies the Huggingface LLM to inference on
• -p_word: Specifies the word embedding to project on. Should be female for GenderBias-QA and republican for PoliticBias-QA

If you find this work useful, please cite our paper:

@misc{kim2025mitigatingbiasragcontrolling,
      title={Mitigating Bias in RAG: Controlling the Embedder}, 
      author={Taeyoun Kim and Jacob Springer and Aditi Raghunathan and Maarten Sap},
      year={2025},
      eprint={2502.17390},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.17390}, 
}