Bot or Not: How well can small LMs behave as bots compared to Large LMs?

In the ever-evolving race between social media bots and detectors, one crucial question remains: can detectors reliably distinguish between human-written content and bot-generated text? If bots can deceive detectors, there is a significant risk of spreading false or misleading information. In this study, we simulate this competition by testing our small language model, Phi-3.5-mini-instruct, combined with techniques such as instruction tuning, Retrieval-Augmented Generation (RAG), and post-processing, can generate tweets that are indistinguishable from human-written content. We challenge this against a much capable model, GPT-4, acting as the detector. If the Small Language Model (SLM) can deceive the much larger one, it would not only prove the prevalence of low-cost undetected social media bots, but it could also provide insights into improving detection systems based on vulnerabilities exposed by this attack. Our research aims to bridge the gap between model size and performance, testing whether smaller, cost-effective models can rival their larger counterparts in generating high-quality, human-like text. This study could shed light on the “Bot or Not” dilemma, offering novel solutions for social media in the ongoing battle against bots.

Retrieval-Augmented Generation with Wikipedia and Breaking news

RAG plays a key role in improving the relevance of tweet generation. We build a vector database using Wikipedia as the foundation documents, and to keep the model informed with the latest news, we scrape articles daily from various sources. As trending tweet topics change daily, we aim to explore whether RAG can help the model integrate up-to-date information and generate tweets that are not only human-like but also more relevant and realistic to confuse the detector further, particularly when addressing specific topics or breaking news.

Project Poster

Below is the poster summarizing our project and key findings:

Specification of dependencies

Install the libraries:

pip install -r requirements.txt

Training code

Notebooks inside the /model folder.

Evaluation code

Detector.ipynb.

Analysis code

Notebooks inside the /analysis folder.

Project Folder Structure

The project is organized into the following structure:

project_root/
├── [analysis/](./analysis)           # Contains the analysis results.
├── [data/](./data)               # Handles tweet data cleaning and preprocessing.
├── [eval/](./eval)               # Includes the tweet generation results for each version of the model.
├── [graphs/](./graphs)             # Stores analysis graphs for better visualization.
├── [model/](./model)              # Contains approaches and implementations for training the tweet bot.
├── [poster/](./poster)             # Project poster materials.
└── [Detector.ipynb](./Detector.ipynb)      # Notebook for defining the GPT-4o Bot Detector and setting up the competition environment for Phase 1 and Phase 2.

Pre-trained models

Two tuned models released in Huggingface.

• Fine-Tuned with 100k Broad Filtering Tweets:

  • Model: https://huggingface.co/AlanYky/phi-3.5_tweets_instruct
  • Data: https://huggingface.co/datasets/AlanYky/tweets_instruct_100k_1
  • This dataset comprises 100,000 tweets filtered using broad criteria to maintain general relevance and clarity. The filtering steps include:
    • Removing mentions (@user).
    • Excluding tweets that are too short or too long.
    • Filtering out tweets with excessive word repetition.

• Fine-Tuned with 50k High-Quality Filtering Tweets:

  • Model: https://huggingface.co/AlanYky/phi-3.5_tweets_instruct_50k
  • Data: https://huggingface.co/datasets/AlanYky/tweets_instruct_v2
  • This dataset includes 50,000 tweets selected through rigorous high-quality filtering. Built on top of broad filtering, the tweets have been further refined by:
    • Removing excessive emojis.
    • Excluding tweets with links or excessive symbols like hashtags.
    • Filtering out tweets containing advertisement-related words.
    • Excluding retweets (tweets containing 'RT').

How to reproduce the results

1. Run /model/tweets-instruct-tuning/tweet_instruct_50k.ipynb and model/tweets-instruct-tuning/tweet_instruct_100k.ipynb to reproduce two models.
2. Run /Detector.ipynb to reproduce candidates model performance comparison.
3. Run /model/lexical-distribution-post-processing/post_processing_human_like_selection.ipynb to reproduce post-processing pipeline.
4. Run all notebooks inside /analysis folder to reproduce the analysis for the generated tweets.

Acknowledgement

Thanks https://newscatcherapi.com/ for free news access.