IC-1149ChatGPT and Llama-2-7b-chat fail to recognize unanswerable questions on SQuAD 2.0, with Llama-2-7b-chat scoring only 3.72% accuracy on no-answer questions

Linyi Yang, Shuibai Zhang, Zhuohao Yu, Guangsheng Bao, Yidong Wang, Jindong Wang, Ruochen Xu, Wei Ye, Xing Xie, Weizhu Chen, Yue Zhang

SourceSupervised Knowledge Makes Large Language Models Better In-context Learners

On the SQuAD 2.0 hold-out test, the paper measures how well each model handles unanswerable questions (questions with no valid answer in the context). ChatGPT achieves 32.35% accuracy on no-answer questions, while Llama-2-7b-chat achieves only 3.72%, meaning it almost never outputs the correct 'I don't know' response. The authors note that Llama-2-7b-chat 'struggled with format adherence and hallucinations, especially in answering no-answer questions.' In-context learning with mixed no-answer and has-answer examples improves no-answer handling but reduces has-answer accuracy.

Evidence
correlational
Key metric
ChatGPT acc.no. = 32.35%, Llama-2-7b-chat acc.no. = 3.72%; ChatGPT acc.has. = 79.35%, Llama-2-7b-chat acc.has. = 31.87%
Caveat
The SQuAD 2.0 evaluation uses a hold-out test where in-context examples are drawn from the training set; the prompt length is controlled under 4,096 tokens for Llama-2.
Model
ChatGPT, Llama 2 / Llama 2 base
Concepts
Failure mode
Datasets
SQuAD 2.0 [eval]
Related findings
IC-1148, IC-1150
Extraction
automatic-extraction