Challenges and Considerations in Language Model Evaluation
summary
Evaluation shows tiny prompt edits can flip scores from near 0% to high values, for example a trailing space or whitespace changes accuracy; few‑shot example ordering and prompt formatting shift performance and at ~500B training tokens models switch preference between closed‑form and multiple‑choice. Benchmarks suffer dataset contamination and saturation—GSM‑8K saturation and GSM‑1K underperformance reveal overfitting—and 95% vs 96% is often meaningless. Practical fixes: publish evaluation code, release prompts and model outputs, and use tools like Pythia, EleutherAI’s LM Evaluation Harness, Helm, OpenCompass, or InspectAI.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:25A Key Challenge in LM Evaluation
- 1:22What do we want to evaluate?
- 4:10LM - Specific Complications
- 9:20Evaluating Models vs Systems
- 11:59Life of a Benchmark
- 14:50Overfitting
- 22:09Addressing Evaluation Pitfalls
- 23:53LM Evaluation is Challenging
related talks
14mImproving Large Language Model Performance
54mKeynote: Can you trust your (large language) model?
16mBest Practices for Open Multilingual LLM Evaluation
16mNavigating the Uncharted: Ensuring Prompt Quality in the Age of Language Models
01h 06mThe Future of Language Models: A Perspective on Evaluation
01h 04mCan you trust your (large language) model?
20mFive hard earned lessons about Evals
54mKeynote: Can you trust your (large language) model?