Checkpoint doesn't seem to match hosted nanochat app

#6
by rfernand - opened

When I ask this checkpoint (using chat_web.py) who created it, I get very different answer from the app hosted at: https://nanochat.karpathy.ai/. These checkpoint doesn't seem to have been trained with the same "identity_conversations.jsonl" file that was specified for the nanochat Github repo. Can you explain how these checkpoints are different?

Also I notice that the hosted app has streaming enabled, but chat_web.py doesn't (when hosted on Windows, output is only visible after it all has been generated). Any idea why?

Thanks.

rfernand changed discussion title from Which training stage does this checkpoint represent? to Checkpoint doesn't seem to match hosted nanochat app

Sign up or log in to comment