Fingerprint for LLM

This repository is a simple implementation of fingerprint for LLM.

Code for paper A Fingerprint for Large Language Models

Usage

Run load_model.py directly.

The format of config.json in dir model is like:

[
  {
    "name": "gemma",
    "version": "2b",
    "new_version": "2b-1.1-it",
    "standard": "<path of the oracle model>",
    "path": {
      "dir": "<path of dir for fine-tuned model>",
      "sub": "results"
    },
    "dataset": ["samsum", "tatsu-lab"],
    "huggingface": true
  }
]

Requirements

torch
transformers
datasets
peft