GDRT

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 23 Open GitHub ↗

[WWW2026] The official code for paper "Does LLM Focus on the Right Words? Mitigating Context Bias in LLM-based Recommenders"

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Does LLM Focus on the Right Words? Mitigating Context Bias in LLM-based Recommenders

![Static Badge](https://img.shields.io/badge/Conference-WWW2026-FF8C00?label=Conference)

This is the Official implementation for our WWW 2026 paper.

> Bohao Wang, Jiawei Chen, Feng Liu, Changwang Zhang, Jun Wang, Canghong Jin, Chun Chen, Can Wang 2026. Does LLM Focus on the Right Words? Mitigating Context Bias in LLM-based Recommenders. [arXiv link](https://arxiv.org/abs/2510.10978)

## Environment

- torch==2.6.0
- transformers==4.49.0
- marisa-trie==1.2.1
- [genre](https://github.com/facebookresearch/GENRE)

## Directory Structure

- `train_gdrt.py`: Code for training the GDRT model.
- `train_sft.py`: Code for training the SFT model.
- `inference_evaluate.py`: Code for running inference and evaluating model performance.

## Scripts

### Training

- To train the GDRT model:
  `group_num` represents the number of groups, and `dro_tau` represents the parameter  $\tau$.

```
  CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch train_gdrt.py --dataset_name Toy --sample 10000 --group_num 2 --dro_tau 0.2
  ```

```
  CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch train_gdrt.py --dataset_name Clothing --sample 10000 --group_num 5 --dro_tau 0.2
  ```

- To train the SFT model:

```
  CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch train_sft.py --dataset_name Toy --sample 10000
  ```

```
  CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch train_sft.py --dataset_name Clothing --sample 10000
  ```

### Evaluate

- Run inference and evaluation:

```
  CUDA_VISIBLE_DEVICES=0,1,2,3 accelerate launch inference_evaluate.py  --lora_weights_father_path "./save_GDRT_model_Llama3.2/Toy/sample10000_epoch5/0"
  ```

Replace `lora_weights_father_path` with the path to the parent directory of the model checkpoint.

## Citation

If you find the paper useful in your research, please consider citing:

```
@misc{wang2026doesllmfocusright,
      title={Does LLM Focus on the Right Words? Mitigating Context Bias in LLM-based Recommenders}, 
      author={Bohao Wang and Jiawei Chen and Feng Liu and Changwang Zhang and Jun Wang and Canghong Jin and Chun Chen and Can Wang},
      year={2026},
      eprint={2510.10978},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2510.10978}, 
}
```

GDRT

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README