IProX

Project Introduction

IProX (Influence-Preserving Proxies) is a two-stage framework for scalable gradient-based data selection in LLM fine-tuning.

Gradient-based selectors (for example, TracIn and influence-function variants) can improve supervised fine-tuning data quality, but they are expensive on large models. IProX addresses this by constructing a smaller proxy directly from a target LLM while preserving influence-related behavior. This repository currently uses a TracIn-based implementation for gradient-based influence estimation.

The pipeline has two stages:

1. Influence-Preserving SVD (IPSVD): compresses target linear layers into low-rank factors with influence-aware initialization.
2. Gradient Alignment: aligns proxy gradients with target gradients in factor space, and anchors proxy logits for stability.

This repository provides code for proxy initialization and training with gradient alignment.

Environment Setup and Training

1. Create and activate an environment

cd IProX
conda create -y -n Iprox python=3.10
conda activate Iprox

2. Install dependencies

Install PyTorch first, then install project dependencies:

pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu126
pip install -r requirements.txt

3. Run training

Example command:

python init_iprox.py \
  --model_name_or_path Qwen/Qwen2.5-0.5B \
  --train_files data/train/processed/dolly/dolly_data.jsonl \
  --sparsity 0.5 \
  --init_method IPSVD \
  --epochs 5 \
  --batch_size 4 \
  --output_dir ../models

Directory Acquisition

1. Repository structure

IProX/
├── init_iprox.py
├── requirements.txt
├── data/
│   └── train/processed/dolly/dolly_data.jsonl
└── utils/
    ├── get_training_dataset.py
    ├── init_with_ipsvd.py
    ├── grad_align.py
    └── util.py

2. Expected data location

The default training file path is:

data/train/processed/dolly/dolly_data.jsonl

If you use your own data, place it anywhere accessible and pass it with --train_files.

Citation Format

If you use this codebase, please cite:

@inproceedings{chen2026influencepreserving,
  title={Influence-Preserving Proxies for Gradient-Based Data Selection in {LLM} FineTuning},
  author={Sirui Chen and Yunzhe Qi and Mengting Ai and Yifan Sun and Ruizhong Qiu and Jiaru Zou and Jingrui He},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=PDNpRLxDlI}
}