We Poisoned an LLM’s Training Data. Here’s What Broke (and What Didn’t).

1. 1Corrupt 25% of human feedback data to test an LLM's robustness
2. 2Poison 100% of the training data to observe the model's behavior
3. 3Analyze the results to identify potential vulnerabilities in the model's safety guardrails
4. 4Implement data validation and verification techniques to prevent data poisoning
5. 5Test the model's performance on a separate, clean dataset to evaluate its robustness

ML engineers and researchers can benefit from understanding the vulnerabilities of LLMs to data poisoning, while data scientists and product managers should be aware of the potential risks to AI model safety

💡 Data poisoning can have significant effects on LLM safety, even with partial corruption

🚨 Corrupting just 25% of human feedback can silently strip an LLM's safety guardrails! 🤖