github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[PR #15627] fix: preserve strip_accents preprocessing in BERT tokenizer conversion #25728

New Issue
Open
opened 2026-04-19 18:24:42 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-19 18:24:42 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/15627
Author: @MasterOfFeelingFish
Created: 4/16/2026
Status: 🔄 Open

Base: main ← Head: clawoss/fix/bert-strip-accents

📝 Commits (1)

  • bc4fa62 fix: preserve strip_accents preprocessing in BERT tokenizer conversion

📊 Changes

7 files changed (+43 additions, -9 deletions)

View changed files

📝 convert/convert_bert.go (+1 -0)
📝 convert/convert_nomicbert.go (+1 -0)
📝 convert/tokenizer.go (+11 -2)
📝 model/models/bert/embed.go (+1 -1)
📝 model/models/nomicbert/model.go (+1 -0)
📝 tokenizer/wordpiece.go (+26 -5)
📝 tokenizer/wordpiece_test.go (+2 -1)

📄 Description

Summary

BERT-derived embedding models produce incorrect embeddings for non-ASCII text because the strip_accents preprocessing step is dropped during GGUF conversion.

Root Cause

The HuggingFace BasicTokenizer used by BERT models applies NFD normalization and strips combining diacritical marks (accents) before WordPiece tokenization. This preprocessing is essential - without it, words like Hokkaidō and Éire fail to match their ASCII equivalents (Hokkaido and Eire) because the diacritics cause the words to tokenize to [UNK].

Changes

  1. convert/tokenizer.go: Read strip_accents field from okenizer_config.json during tokenizer parsing
  2. convert/convert_bert.go & convert/convert_nomicbert.go: Store strip_accents in GGUF metadata
  3. tokenizer/wordpiece.go:
    • Add stripAccents field to WordPiece struct
    • Add stripAccents() helper function to remove combining diacritical marks (U+0300-U+036F)
    • Apply accent stripping in Encode() when stripAccents is enabled
  4. model/models/bert/embed.go & model/models/nomicbert/model.go: Pass strip_accents config to NewWordPiece()

Impact

This fix ensures that models like mxbai-embed-large,
omic-embed-text, and ll-minilm correctly handle non-ASCII text, matching the behavior of the original HuggingFace models.

Fixes #15609

🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/ollama/ollama/pull/15627 **Author:** [@MasterOfFeelingFish](https://github.com/MasterOfFeelingFish) **Created:** 4/16/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `clawoss/fix/bert-strip-accents` --- ### 📝 Commits (1) - [`bc4fa62`](https://github.com/ollama/ollama/commit/bc4fa6290d858eccdccc5a54fcbbea256a799a59) fix: preserve strip_accents preprocessing in BERT tokenizer conversion ### 📊 Changes **7 files changed** (+43 additions, -9 deletions) <details> <summary>View changed files</summary> 📝 `convert/convert_bert.go` (+1 -0) 📝 `convert/convert_nomicbert.go` (+1 -0) 📝 `convert/tokenizer.go` (+11 -2) 📝 `model/models/bert/embed.go` (+1 -1) 📝 `model/models/nomicbert/model.go` (+1 -0) 📝 `tokenizer/wordpiece.go` (+26 -5) 📝 `tokenizer/wordpiece_test.go` (+2 -1) </details> ### 📄 Description ## Summary BERT-derived embedding models produce incorrect embeddings for non-ASCII text because the strip_accents preprocessing step is dropped during GGUF conversion. ### Root Cause The HuggingFace BasicTokenizer used by BERT models applies NFD normalization and strips combining diacritical marks (accents) before WordPiece tokenization. This preprocessing is essential - without it, words like Hokkaidō and Éire fail to match their ASCII equivalents (Hokkaido and Eire) because the diacritics cause the words to tokenize to [UNK]. ### Changes 1. **convert/tokenizer.go**: Read strip_accents field from okenizer_config.json during tokenizer parsing 2. **convert/convert_bert.go** & **convert/convert_nomicbert.go**: Store strip_accents in GGUF metadata 3. **tokenizer/wordpiece.go**: - Add stripAccents field to WordPiece struct - Add stripAccents() helper function to remove combining diacritical marks (U+0300-U+036F) - Apply accent stripping in Encode() when stripAccents is enabled 4. **model/models/bert/embed.go** & **model/models/nomicbert/model.go**: Pass strip_accents config to NewWordPiece() ### Impact This fix ensures that models like mxbai-embed-large, omic-embed-text, and ll-minilm correctly handle non-ASCII text, matching the behavior of the original HuggingFace models. Fixes #15609 --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
GiteaMirror added the pull-request label 2026-04-19 18:24:43 -05:00
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label pull-request
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#25728
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?