Spaces:
Running
on
CPU Upgrade
Dedup redundant entries (for example we have 3 entries for llama-65 right now)
Right now, the dashboard is listing 3 entries that have the exact same score on each benchmark:
I suspect those entries each point to the same underlying set of weights and config, and the leaderboard would benefit from having some mechanism for consolidating these (and other similar instances as they appear) into a single entry.
Alas I don't think identical scores and similar names are sufficient to consistently flag true duplicates. For example, we also have:
Yet one of those models describes itself as a "Merge of SuperHOT-LoRA-prototype and llama-30b", and so presumably has different weights than a pure llama-30b model.
Maybe tracking a checksum of relevant model files for each entry would help to flag clear duplicates?
Cool idea!
Towards the bottom there are also several delta models that have been uploaded without being merged. They definitely need to be either retested or removed.
Hi! We added the model sha to the ExtendedView, in order to get more information on the models and their weights