3 ms·
They don't mention what datasets were used. I've come across too many models in the past which gave amazing results because benchmarks leaked into their trainin
by nabakin 3y ago
They don't mention what datasets were used. I've come across too many models in the past which gave amazing results because benchmarks leaked into their training data. How are we supposed to verify one of these HuggingFace datasets didn't leak the benchmarks into the training data boosting their results? Did they do any checking of their datasets for leaks? How are we supposed to know this is a legit result?
At this point, it should be standard practice to address this concern. Any model which fails provide good evidence they don't have benchmark leaks, should not be trusted until its datasets can be verified, the methodology can be replicated, or a good, independent, private benchmark can be made and can be used to evaluate the model.
- londons_explore 3y agoOne solution is to come up with a new benchmark yourself. Manually benchmarking it by coming up with 20 questions and feeding it to a pair of models and blindly choosing the best result can give you a pretty good figure. And that can probably be done in under 20 mins of human time.
- poser-boy 3y agoThis is frankly the best benchmark you can use on an LLM.
- avereveard 3y agoYou can ask gpt4 or other high value model to rate two chat logs for coherency etc, not as accurate as human evaluation, but you don't have to read thousand lines of text if comparing many models.
- brucethemoose2 3y agoThis is problematic if you are comparing a model in the same base family as the evaluator, as it will probably favor itself because it literally has the sequences it would naturally emit.
- Breza 3y agoEspecially since most devs have a specific use case in mind. Coming up with tests that are tailored to your needs will always be more informative than off the shelf metrics.
- mufasachan 3y agoOne explaination https://x.com/yampeleg/status/1707127722743325106?s=46&t=CxaRJxotO6kj3op6jNibrw https://x.com/yampeleg/status/1707127722743325106?s=46&t=Cxa... I would be curious what does he mean by "semi-automated system for detecting benchmark leaks. " though.
- brucethemoose2 3y agoAFAIK, such tests just feed the model chopped up bits of the evaluation data as raw strings with zero temperature. If it completes them verbatim, its probably in the training dataset.