Fill-Mask
Transformers
Safetensors
modernbert
masked-lm
long-context
timpal0l commited on
Commit
9c43b95
·
verified ·
1 Parent(s): 9228173

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +29 -1
README.md CHANGED
@@ -64,7 +64,35 @@ See training details [here](https://github.com/timpal0l/ModernBERT/blob/main/tra
64
  ```
65
  ## Intended Use
66
  * Fill-mask inference, embedding extraction and fine-tuning for Scandinavian downstream NLP tasks (classification, NER, QA, etc.).
67
- * Drop-in replacement for BERT-style encoders (omit `token_type_ids`).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
68
  ## Limitations & Biases
69
  * Web corpora can contain noise, stereotypes and sensitive content despite filtering.
70
  * RoPE extrapolation beyond 8 k tokens is untested and may degrade.
 
64
  ```
65
  ## Intended Use
66
  * Fill-mask inference, embedding extraction and fine-tuning for Scandinavian downstream NLP tasks (classification, NER, QA, etc.).
67
+ * Drop-in replacement for BERT-style encoders (omit `token_type_ids`).
68
+ ## Fill-mask
69
+ ```python
70
+ from transformers import pipeline
71
+ unmasker = pipeline('fill-mask', model='AI-Sweden-Models/ModernBERT-large')
72
+ unmasker("Huvudstaden i Sverige är [MASK].")
73
+ ```
74
+ ```python
75
+ [{'score': 0.5732529759407043,
76
+ 'token': 2961,
77
+ 'token_str': ' Stockholm',
78
+ 'sequence': 'Huvudstaden i Sverige är Stockholm.'},
79
+ {'score': 0.06222670152783394,
80
+ 'token': 4481,
81
+ 'token_str': ' Göteborg',
82
+ 'sequence': 'Huvudstaden i Sverige är Göteborg.'},
83
+ {'score': 0.02539575845003128,
84
+ 'token': 5882,
85
+ 'token_str': ' Malmö',
86
+ 'sequence': 'Huvudstaden i Sverige är Malmö.'},
87
+ {'score': 0.024683712050318718,
88
+ 'token': 19931,
89
+ 'token_str': ' Norrköping',
90
+ 'sequence': 'Huvudstaden i Sverige är Norrköping.'},
91
+ {'score': 0.02418600209057331,
92
+ 'token': 28202,
93
+ 'token_str': ' Solna',
94
+ 'sequence': 'Huvudstaden i Sverige är Solna.'}]
95
+ ```
96
  ## Limitations & Biases
97
  * Web corpora can contain noise, stereotypes and sensitive content despite filtering.
98
  * RoPE extrapolation beyond 8 k tokens is untested and may degrade.