Gemma 2 / Mistral 7b v.1 issues

Open
#14 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
25/100
Issue type
Bug
Clarity
Needs clarification
Activity status
Stale
Tech stack
python

Research direction

Start with the zero-shot evaluation flow and the compute_universal_function_vector function, then reproduce the top-1 accuracy results for Gemma 2 and Mistral 7B v0.1 across layers. Compare their indirect-effect aggregation and top-head selection with models that work; done means identifying and documenting the model-specific cause or providing a verified fix.

Written by the indexing model from the issue text.

Description

Hi, I have been running the zero-shot evaluation on several models. For models that don't already have a universal top head list, I compute the indirect effect scores for each abstractive task, and, for the tasks with 10-shot performance greater than the majority baseline, aggregate them. I then add the top k heads (k is roughly proportional to # heads in the model) list to the compute_universal_function_vector function. I then evaluate using these top heads. For most models, I seem to get good results. However, for gemma 2 (of all sizes) and Mistral 7b v.01, I almost always get a top-1 accuracy lower than 10%, with a score that doesn't change much between the layers. Do you know why the tests don't work with these models?
Thanks!

Dominant language
Python
Stars
200
Forks
47
PR merge metrics
No merged PRs in 30d

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from ericwtodd/function_vectors

All issues in ericwtodd/function_vectors

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.