Providing byte level offsets for effective alignment in Cross-Tokenizer On-Policy Distillation #1880
+62
−18
Conversation
This file contains hidden or 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
There was a problem hiding this comment.
Pull Request Overview
This PR adds an offset_type parameter to the encode and encode_batch methods, allowing users to choose between character-based offsets ("char"), byte-based offsets ("byte"), or no offsets ("none") for faster encoding. The default is "char" to maintain backward compatibility.
- Adds
offset_typeparameter to both Rust and Python encoding methods - Routes to appropriate internal methods based on offset type selection
- Provides input validation with helpful error messages
Reviewed Changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
| bindings/python/src/tokenizer.rs | Adds offset_type parameter to encode and encode_batch Rust methods with validation and routing logic |
| bindings/python/py_src/tokenizers/implementations/base_tokenizer.py | Adds offset_type parameter to Python wrapper methods with documentation |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| .into_pyerr::<exceptions::PyValueError>()) | ||
| } | ||
| }; | ||
|
|
There was a problem hiding this comment.
Trailing whitespace should be removed from this empty line for consistency with code style.
Suggested change
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
issue
#1881
Our team tried different aligning solutions when implementing our self-developed On-Policy distillation method, including trl's existing implementation (which repeatedly calls the decode() method but encounters correctness issues in special cases and has high computational overhead).
Later we found a better method: by making one call to the encode() method to get byte-level offsets for all tokens, we can effectively avoid BPE's complexity, and byte level offsets are also compatible with all other types of tokenizers. Additionally, for distillation between two BPE tokenizers, we can get more accurate alignment by skipping string as an intermediate modality.
Therefore, we hope to merge this simple patch to expose the byte-level offset calculation already supported in the Rust code for use by Python classes.
More description at:
huggingface/trl#4393