PAIR-code / PAIR-code/tiny-transformers
Fix scaling in attention computation
Open
Nobody has claimed this yet.
- Dominant language
- Jupyter Notebook
- Stars
- 24
- Forks
- 6
- PR merge metrics
- No merged PRs in 30d
Description
Currently we are not dividing the raw attention by the square root of dimension, when we should.
It's possible that this explains some of the losses peaks during training of tiny models.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
No file or test is named. Locate the attention computation in the notebook or source, identify the raw attention and its dimension, and inspect how training loss is measured. Done means the raw attention is scaled by the square root of its dimension and the existing notebook or test checks pass.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- jupyter-notebook
- Domain
- machine-learning
- Issue type
- Bug
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 45/100