Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,39 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
datasets:
|
3 |
+
- ascolda/ru_en_Crystallography_and_Spectroscopy
|
4 |
+
language:
|
5 |
+
- ru
|
6 |
+
- en
|
7 |
+
metrics:
|
8 |
+
- bleu
|
9 |
+
pipeline_tag: translation
|
10 |
+
tags:
|
11 |
+
- chemistry
|
12 |
+
---
|
13 |
+
# nllb-200-distilled-600M_ru_en_finetuned_crystallography
|
14 |
+
|
15 |
+
This model is a fine-tuned version of facebook/nllb-200-distilled-600M trained on the ascolda/ru_en_Crystallography_and_Spectroscopy dataset
|
16 |
+
It achieves the following results on the evaluation set:
|
17 |
+
- Loss: 0.5602
|
18 |
+
- Bleu: 56.5855
|
19 |
+
|
20 |
+
## Model description
|
21 |
+
|
22 |
+
The finetuned model yieled better performance on the machine translation task of domain-specific scientific articles related to the Crystallography and Spectroscopy domain.
|
23 |
+
|
24 |
+
## Metrics used to describe the fine-tuning effect
|
25 |
+
|
26 |
+
Below is the comparison of the translation quality metrics for the original NLLB model and my finetuned version. Evaluation is focused on: (1) general translation quality, (2) quality of translation of specific
|
27 |
+
terminology, and (3) uniformity of translation of domain-specific terms in different contexts.
|
28 |
+
|
29 |
+
(1) The general translation quality was evaluated using the Bleu metric.
|
30 |
+
|
31 |
+
(2) Term Success Rate. In the terminology success rate we compared the machine-translated terms with their dictionary equivalents by checking for the presence of the reference terminology translation in the output by the regular expression match.
|
32 |
+
|
33 |
+
(3) Term Consistency. This metric looks at whether technical terms are translated uniformly across the entire text corpus in different contexts. We aim for high consistency,
|
34 |
+
measured by the low occurrence of multiple translations for the same term within the evaluation dataset.
|
35 |
+
|
36 |
+
| Model | BLEU | Term Success Rate | Term Consistency |
|
37 |
+
|:--------------------------------------------------------------:|:-------:|:-------------------:|:----------------:|
|
38 |
+
| nllb-200-distilled-600M | 38.19 | 0.246 | 0.199 |
|
39 |
+
| nllb-200-distilled-600M_ru_en_finetuned_crystallography | 56.59 | 0.573 | 0.740 |
|