facebookresearch / facebookresearch/blt
Current implementation of patcher does not support sequence packing
- Dominant language
- Python
- Stars
- 2.1k
- Forks
- 196
- PR merge metrics
- No merged PRs in 30d
Description
I am right, if i use sequence packing with custom masking (to not look cross document) i need to manually start new patch at BOS token to prevent situation where end of one document and start of next document fall into one patch?
For example: [bos, 1,2,3, eos, bos, 4,5,6,eos]. So, i don't want the situation like [1,0,0,0,0,_0_,1,0,0,0] to happen.
Also, what is the purpose of _include_next_token_ arg in patcher?
Contributor guide
Research direction
The issue names only the patcher and its include_next_token argument; start by locating that implementation and its existing tests or usage. Clarify how sequence packing should handle BOS and EOS boundaries, what include_next_token controls, and consider the work complete when the expected masking and patch boundaries are covered by tests.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 30/100