Hourly scraped provider throughput and latency data for every active model (i.e. models that have seen some amount of token generation in the past week). Each checkpoint is saved as a git commit, similar to other projects like pge-outages and mcbroken-archive.
To more easily perform analysis across multiple checkpoints, I recommend using
the git-history tool. For this
specific repository, the following command can be used to load all checkpoints
into a SQLite database for more robust analysis.
uvx --with 'git-history>=0.8' git-history file \
model-metrics.db model-metrics.json \
--id model --id provider --ignore-duplicate-ids \
--import re --convert "$(cat git-history-convert.py)"Once that's done, you can query model-metrics.db any way you like. For example, to see the history of the metrics for glm-4.7 on Chutes, use this query:
select i.model, i.provider, iv.tokens_per_second, iv.time_to_first_token, c.commit_at
from item i
join item_version iv on iv._item = i._id
join commits c on c.id = iv._commit
where i.model = 'z-ai/glm-4.7' and i.provider = 'Chutes'
order by c.commit_at;