Train parameters exclusively in specific ranges #1390
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| @@ -16,12 +16,12 @@ output_dir: ./qlora-out | |||
|
|
|||
| ## You can optionally freeze the entire model and unfreeze a subset of parameters | |||
| unfrozen_parameters: | |||
| # - lm_head.* | |||
There was a problem hiding this comment.
The previous example was technically wrong (but worked) because this becomes the following:
lm_head\.* and checks if the dot repeats 0+ times
seungduk-yanolja
added a commit
to Y-IAB/axolotl
that referenced
this pull request
Mar 19, 2024
* Train parameters exclusively in specific ranges * Fix the style and update docs * Update yaml example
8 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Train parameters exclusively in specific ranges
Description
This PR enables the freezing configuration to specify parameter ranges so that we can ignore gradients for specific tokens.
For more details, please take a look at this technical report.
https://arxiv.org/abs/2402.14714
Motivation and Context
When some tokens are added to the vocab, we may want to train only those specific tokens. For example, while fine-tuning, imagine you want to train specific control tokens only.
This technique can be also used when expanding the vocabulary for non-English languages.
How has this been tested?
I wrote unit tests for it.
Screenshots (if appropriate)
N/A
Types of changes
New feature (non-breaking change which adds functionality)