README.md
| 1 | --- |
| 2 | pipeline_tag: text-classification |
| 3 | --- |
| 4 | |
| 5 | <br> |
| 6 | |
| 7 | # RADAR Model Card |
| 8 | |
| 9 | ## Model Details |
| 10 | |
| 11 | RADAR-Vicuna-7B is an AI-text detector trained via adversarial learning between the detector and a paraphraser on human-text corpus ([OpenWebText](https://huggingface.co/datasets/Skylion007/openwebtext)) and AI-text corpus generated |
| 12 | based on [OpenWebText](https://huggingface.co/datasets/Skylion007/openwebtext). |
| 13 | |
| 14 | - **Developed by:** [TrustSafeAI](https://huggingface.co/TrustSafeAI) |
| 15 | - **Model type:** An encoder-only language model based on the transformer architecture (RoBERTa). |
| 16 | - **License:** [Non-commercial license](https://huggingface.co/lmsys/vicuna-7b-v1.1#model-details) (inherited from Vicuna-7B-v1.1) |
| 17 | - **Trained from model:** [RoBERTa](https://arxiv.org/abs/1907.11692) |
| 18 | |
| 19 | |
| 20 | ### Model Sources |
| 21 | |
| 22 | - **Project Page:** https://radar.vizhub.ai/ |
| 23 | - **Paper:** https://arxiv.org/abs/2307.03838 |
| 24 | - **IBM Blog Post:** https://research.ibm.com/blog/AI-forensics-attribution |
| 25 | |
| 26 | ## Uses |
| 27 | Users could use this detector to assist them in detecting text generated by large language models. |
| 28 | Please note that this detector is trained on AI-text generated by Vicuna-7B-v1.1. As the model only supports [non-commercial use](https://huggingface.co/lmsys/vicuna-7b-v1.1#model-details), the intended users are **not allowed to involve this detector into commercial activities**. |
| 29 | |
| 30 | ## Get Started with the Model |
| 31 | Please refer to the following guidelines to see how to locally run the downloaded model or use our API service hosted on Huggingface Space. |
| 32 | - Google Colab Demo: https://colab.research.google.com/drive/1r7mLEfVynChUUgIfw1r4WZyh9b0QBQdo?usp=sharing |
| 33 | - Huggingface API Documentation: https://trustsafeai-radar-ai-text-detector.hf.space/?view=api |
| 34 | |
| 35 | ## Training Pipeline |
| 36 | |
| 37 | We propose adversarial learning between a paraphraser and our detector. The paraphraser's goal is to make the AI-generated text more like human-writen and the detector's goal is to |
| 38 | promote it's ability to identify the AI-text. |
| 39 | |
| 40 | - **(Step 1) Training Data preparation**: Before training, we use Vicuna-7B to generate AI-text by performing text completion based on the prefix span of human-text in [OpenWebText](https://huggingface.co/datasets/Skylion007/openwebtext). |
| 41 | |
| 42 | - **(Step 2) Update the paraphraser** During training, the paraphraser will do paraphrasing on the AI-text generated in **Step 1**. And then collect the reward returned by the detector to update the paraphraser using Proxy Proximal Optimization loss. |
| 43 | |
| 44 | - **(Step 3) Update the detector** The detector is optimized using the logistic loss on the human-text, AI-text and paraphrased AI-text. |
| 45 | |
| 46 | See more details in Sections 3 and 4 of this [paper](https://arxiv.org/pdf/2307.03838.pdf). |
| 47 | |
| 48 | ## Ethical Considerations |
| 49 | We suggest users use our tool to assist with identifying AI-written content at scale and with discretion. If the detection result is to be used as evidence, further validation steps |
| 50 | are necessary as RADAR cannot always make correct predictions. |