Datasets:

mteb

/

results

like 6

Hi! I'm trying to upload results on LongEmbed. As an attempt, I have only uploaded that of e5-mistral up until now. There is one thing I'm not quite sure about, regarding the passkey and needle task. While for the other four tasks, we simply report the ndcg@10 score of test split, these 2 tasks have many splits ({test_256, test_512, test_1024, test_2048, test_4096, test_8192, test_16384, test_32768}) , and we average over the ndcg@1 score of each split to represent the final score on these two tasks. In this case, do I need to do something or the mteb leaderboard will handle it automatically?

I think the result files are good as is, but we will need some code to average them either in the results.py here or in the leaderboard. I think it is best to do it in the results.py file but I may be wrong.

Do you want to add the other result files? Then I can merge this & take care of adding the leaderboard tab etc :)

yes, sorry for the late reply! I will add other result files recently :-)

Hi @Muennighoff , I've added more result files. Really appreciate your support!

Some results are still missing no? cc @dwzhu E.g. OpenAI etc.

I've added a preliminary leaderboard tab here: https://huggingface.co/spaces/mteb/leaderboard?task=retrieval&language=longembed

Feel free to edit the tab / add more results. It seems like they don't quite match your paper yet, maybe you could investigate why? (It's probably the metrics i.e. averaging ndcg_at_10 - maybe we need to allow multiple main metrics for a single leaderboard tab; feel free to open a PR on the leaderboard code if that is the case) 🙌

Some results are still missing no? cc @dwzhu E.g. OpenAI etc.

I will add other results gradually, including the latest OpenAI models that have not been reported in our paper.

I've added a preliminary leaderboard tab here: https://huggingface.co/spaces/mteb/leaderboard?task=retrieval&language=longembed

Feel free to edit the tab / add more results. It seems like they don't quite match your paper yet, maybe you could investigate why? (It's probably the metrics i.e. averaging ndcg_at_10 - maybe we need to allow multiple main metrics for a single leaderboard tab; feel free to open a PR on the leaderboard code if that is the case) 🙌

Really appreciate your generous help! I guess it is because the scores we report for Passkey and Needle are Acc@1 (or nDCG@1, they are all the same), but the scores in leaderboard are nDCG@10. @Muennighoff

I would also like to ask, what kind of modifications do I need to make, so that I can get the results on leaderboard right? Is it just modify this: https://github.com/embeddings-benchmark/mteb/blob/fe912bc2d676772ce059522639bcd14e9c80ecb2/mteb/tasks/Retrieval/eng/LEMBPasskeyRetrieval.py#L33C21-L33C31 ? And, on the leaderboard page, I notice that the Metric is explained as follows: Metric: Normalized Discounted Cumulative Gain @ k (ndcg_at_10). What if I want to explain that for Passkey and Needle, the metric is ndcg_at_1?

I think it requires some changes in the code here: https://huggingface.co/spaces/mteb/leaderboard/blob/main/app.py

ok, I'll try that

That'd be amazing - feel free to open a PR! 🙌
Lmk when the other results are ready so we can merge them in this repo too

Thanks! I have opened a PR here https://huggingface.co/spaces/mteb/leaderboard/discussions/124. I have not managed to run the demo myself due to some networking issues, but I guess the modifications will work. Looking forward to your advice! @Muennighoff