system HF staff commited on
Commit
9fbc81f
1 Parent(s): 41133e8

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +110 -0
README.md ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ ---
3
+ language:
4
+ - en
5
+ thumbnail:
6
+ tags:
7
+ - translation
8
+ - facebook
9
+ - convAI
10
+ license: apache-2.0
11
+ datasets:
12
+ - blended_skill_talk
13
+ metrics:
14
+ - perplexity
15
+ ---
16
+
17
+ # Blenderbot-3B
18
+
19
+ ## Model description
20
+
21
+
22
+ + [Paper](https://arxiv.org/abs/1907.06616).
23
+ + [Original PARLAI Code]
24
+
25
+ The abbreviation FSMT stands for FairSeqMachineTranslation
26
+
27
+ All four models are available:
28
+
29
+ * [wmt19-en-ru](https://huggingface.co/facebook/wmt19-en-ru)
30
+ * [wmt19-ru-en](https://huggingface.co/facebook/wmt19-ru-en)
31
+ * [wmt19-en-de](https://huggingface.co/facebook/wmt19-en-de)
32
+ * [wmt19-de-en](https://huggingface.co/facebook/wmt19-de-en)
33
+
34
+ ## Intended uses & limitations
35
+
36
+ #### How to use
37
+
38
+ ```python
39
+ from transformers.tokenization_fsmt import FSMTTokenizer
40
+ from transformers.modeling_fsmt import FSMTForConditionalGeneration
41
+ mname = "facebook/wmt19-en-ru"
42
+ tokenizer = FSMTTokenizer.from_pretrained(mname)
43
+ model = FSMTForConditionalGeneration.from_pretrained(mname)
44
+
45
+ input = "Machine learning is great, isn't it?"
46
+ input_ids = tokenizer.encode(input, return_tensors="pt")
47
+ outputs = model.generate(input_ids)
48
+ decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
49
+ print(decoded) # Машинное обучение - это здорово, не так ли?
50
+
51
+ ```
52
+
53
+ #### Limitations and bias
54
+
55
+ - The original (and this ported model) doesn't seem to handle well inputs with repeated sub-phrases, [content gets truncated](https://discuss.huggingface.co/t/issues-with-translating-inputs-containing-repeated-phrases/981)
56
+
57
+ ## Training data
58
+
59
+ Pretrained weights were left identical to the original model released by fairseq. For more details, please, see the [paper](https://arxiv.org/abs/1907.06616).
60
+
61
+ ## Eval results
62
+
63
+ pair | fairseq | transformers
64
+ -------|---------|----------
65
+ en-ru | [36.4](http://matrix.statmt.org/matrix/output/1914?run_id=6724) | 33.47
66
+
67
+ The score is slightly below the score reported by `fairseq`, since `transformers`` currently doesn't support:
68
+ - model ensemble, therefore the best performing checkpoint was ported (``model4.pt``).
69
+ - re-ranking
70
+
71
+ The score was calculated using this code:
72
+
73
+ ```bash
74
+ git clone https://github.com/huggingface/transformers
75
+ cd transformers
76
+ export PAIR=en-ru
77
+ export DATA_DIR=data/$PAIR
78
+ export SAVE_DIR=data/$PAIR
79
+ export BS=8
80
+ export NUM_BEAMS=15
81
+ mkdir -p $DATA_DIR
82
+ sacrebleu -t wmt19 -l $PAIR --echo src > $DATA_DIR/val.source
83
+ sacrebleu -t wmt19 -l $PAIR --echo ref > $DATA_DIR/val.target
84
+ echo $PAIR
85
+ PYTHONPATH="src:examples/seq2seq" python examples/seq2seq/run_eval.py facebook/wmt19-$PAIR $DATA_DIR/val.source $SAVE_DIR/test_translations.txt --reference_path $DATA_DIR/val.target --score_path $SAVE_DIR/test_bleu.json --bs $BS --task translation --num_beams $NUM_BEAMS
86
+ ```
87
+ note: fairseq reports using a beam of 50, so you should get a slightly higher score if re-run with `--num_beams 50`.
88
+
89
+ ## Data Sources
90
+
91
+ - [training, etc.](http://www.statmt.org/wmt19/)
92
+ - [test set](http://matrix.statmt.org/test_sets/newstest2019.tgz?1556572561)
93
+
94
+
95
+ ### BibTeX entry and citation info
96
+
97
+ ```bibtex
98
+ @inproceedings{...,
99
+ year={2020},
100
+ title={Facebook FAIR's WMT19 News Translation Task Submission},
101
+ author={Ng, Nathan and Yee, Kyra and Baevski, Alexei and Ott, Myle and Auli, Michael and Edunov, Sergey},
102
+ booktitle={Proc. of WMT},
103
+ }
104
+ ```
105
+
106
+
107
+ ## TODO
108
+
109
+ - port model ensemble (fairseq uses 4 model checkpoints)
110
+