update teacher checkpointer paths in KD config

[ebsmothers]

Fix checkpoint path in one of our KD configs

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2496

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 3609dba with merge base dab36d2 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[pbontrager]

I'm uncomfortable with this PR because it's making a lot of assumptions about where things are saved during fine-tuning that might break in the future. It also makes the recipe dependent on having run another recipe which is new. Could you add more context on how common it is to finetune the teacher first?

[pbontrager]

Why is epoch 0 the right default here?

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 23.15%. Comparing base (dab36d2) to head (745a5e9).
Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2496   +/-   ##
=======================================
  Coverage   23.15%   23.15%           
=======================================
  Files         379      379           
  Lines       22838    22838           
=======================================
  Hits         5289     5289           
  Misses      17549    17549           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ebsmothers]

@pbontrager these are fair comments. Tbh I'm not sure what the right thing to do here is, but would point to GRPO where we do something pretty similar (though ofc that is just in dev for now). The results are better when the teacher model is finetuned first, as discussed in the blog post. So based on that I claim this is the right thing to do, but understand your point around the usage of epoch_0 being a bit finicky.