Summary of Understanding the Effect Of Noise in Llm Training Data with Algorithmic Chains Of Thought, by Alex Havrilla et al.
Understanding the Effect of Noise in LLM Training Data with Algorithmic Chains of Thought
by Alex Havrilla, Maia Iyer
First submitted to arxiv on: 6 Feb 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Large Language Models (LLMs) are trained on massive amounts of text data, but it’s unclear how varying levels of noise affect their performance. This paper investigates the impact of noise on chain-of-thought (CoT) in algorithmically solvable tasks. The authors develop a Traced Integer (TInt) framework to generate customizable noisy execution traces for arithmetic functions and define two types of noise: static, which is applied after CoT computation, and dynamic, which propagates errors during computation. They evaluate the performance of pretrained models prompted and fine-tuned on noised datasets with varying levels of contamination and intensity. The results show that fine-tuned models are robust to high static noise but struggle with low dynamic noise, while few-shot prompted models are more sensitive to even static noise. This research highlights the importance of removing samples containing destructive dynamic noise. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models (LLMs) need huge amounts of text data to learn. But what happens when this data is not perfect? This paper looks at how imperfect data affects how well LLMs do their job. They create a special way to make noisy data and test how well different types of noise affect the models. They find that some models are really good at dealing with messy data, but others get confused easily. The researchers think this is important for making sure we use the right data when training these models. |
Keywords
* Artificial intelligence * Few shot
