This is in line with my own personal experience with LLMs and non-trivial questions. They’re excellent when answering questions on topics you know nothing about, and somehow embarrassingly wrong when you actually know the answer yourself…
It’s not clear to me why we’re still trying to encode all of human knowledge in a single model, instead of teaching the model how to look for answers from an external source (e.g. RAG).
"The benchmark includes questions where at least one LLM confabulated, in order to minimize the number of questions requiring human assessment. Because of this, and since the questions are intentionally adversarial, the absolute percentage should not be used to infer that LLMs frequently confabulate. This leaderboard does not reflect a "typical" hallucination rate."
> instead of teaching the model how to look for answers from an external source (e.g. RAG)
My benchmark specifically focuses on the RAG use case. Even with provided texts, current models still hallucinate.
I stopped playing with larger models and have been pushing smaller models with this improvised system prompt and getting good results. It seems like it forces the model to do multiple passes before giving you any response.
My smaller local models give me less hallucinations than Meta.ai, for example, which generally spits out pleasing answers almost immediately (which are often hallucinations, since I don’t think it is system prompted to be adversarial to the user, or itself). I don’t have the same hallucination issue with Llama3 - 8b locally because of custom system prompts.
The model has all the correct information, so it almost needs to do RAG on itself. Multiple passes on itself seems like a way to do it.
• LLMs, at least GPT models, tend to overstate their confidence. • A frequency-based approach appears to achieve calibration closer to the ideal.
This kinda passes my vibe test. That said, I wonder—rather than running 100 trials, could we approximate this by using something like a log-probability ratio? This would especially apply in cases where answers are yes or no, assuming the output spans more than one token.
> SimpleQA was created to be a greater challenge for frontier models (e.g., GPT-4o scores less than 40%).
"To be included in the dataset, each question had to meet a strict set of criteria: ... and most questions had to induce hallucinations from either GPT-4o or GPT-3.5."
> SimpleQA is a simple but challenging benchmark for evaluating the factuality of frontier models. A main limitation in SimpleQA is its scope—while SimpleQA is accurate it only measures factuality under the constrained setting of short, fact-seeking queries with a single, verifiable answer. Whether the ability to provide factual short answers correlates with the ability to write lengthy responses filled with numerous facts remains an open research question.
OpenAI going to have some rounds of layoffs in the future."I seem, then, in just this little thing to be wiser than this man at any rate; that what I do not know I do not think I know either." - Socratos, from Plato's Apology of Socrates
The steps I took to find this link:
1) Look at simpleqa_eval.py. See that it loads "az://openaipublic/simple-evals/simple_qa_test_set.csv" Hmm, some weird vendored protocol.
2) I don't feel like digging through bf.BlobFile() to figure out how it downloads files and I certainly don't want to generate an API key. Cross fingers and do a Bing web search for "az://openaipublic"
3) That leads me to https://stackoverflow.com/questions/76106366/how-to-use-tikt... Ah ha, this answer has the link https://openaipublic.blob.core.windows.net/encodings/cl100k_... which automatically downloads a file.
4) Poke the relevant parts of the az:// link into this link, and a csv appears.
Why not? Just train an unbelievably gigantic LLM that encodes all human knowledge. A hundred trillion parameters ought to do it.