aktheroy commited on
Commit
d5d5cb3
·
verified ·
1 Parent(s): 52ace3f

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +148 -3
README.md CHANGED
@@ -1,3 +1,148 @@
1
- ---
2
- license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ language:
4
+ - en
5
+ - hi
6
+ - el
7
+ metrics:
8
+ - bleu
9
+ base_model:
10
+ - facebook/m2m100_418M
11
+ ---
12
+ # Model Card for Multilingual Translation Model
13
+
14
+ ## Model Details
15
+
16
+ ### Model Description
17
+
18
+ This model is a fine-tuned version of `facebook/m2m100_418M` for multilingual translation tasks. It supports English (`en`), Hindi (`hi`), and Greek (`el`) as source and target languages. The model has been specifically optimized to ensure accurate and fluent translations across these languages.
19
+
20
+ - **Developed by:** Arun Kumar Roy
21
+ - **Model type:** Transformer-based sequence-to-sequence model for machine translation
22
+ - **Language(s) (NLP):** English, Hindi, Greek
23
+ - **License:** MIT
24
+ - **Finetuned from model:** `facebook/m2m100_418M`
25
+
26
+ ### Model Sources
27
+
28
+ - **Repository:** [Link to model repository]
29
+ - **Demo [optional]:** [Provide a link if applicable]
30
+
31
+ ## Uses
32
+
33
+ ### Direct Use
34
+
35
+ This model can be directly used for multilingual machine translation tasks in English, Hindi, and Greek. Use cases include:
36
+
37
+ - Document translation
38
+ - Real-time conversational translation
39
+ - Educational tools for language learning
40
+
41
+ ### Downstream Use [optional]
42
+
43
+ The model can be further fine-tuned for domain-specific translation tasks such as medical, legal, or technical documents.
44
+
45
+ ### Out-of-Scope Use
46
+
47
+ The model may not perform well for:
48
+ - Languages other than English, Hindi, and Greek.
49
+ - Highly informal, dialectical, or domain-specific text without additional fine-tuning.
50
+ - Use cases requiring strict grammatical correctness for complex legal or academic content.
51
+
52
+ ## Bias, Risks, and Limitations
53
+
54
+ This model inherits potential biases from its training data, which may include:
55
+ - Gender bias in language representation.
56
+ - Cultural or contextual inaccuracies when translating idiomatic expressions.
57
+
58
+ ### Recommendations
59
+
60
+ Users should:
61
+ - Review translations for critical applications.
62
+ - Be cautious when using the model for sensitive or culturally nuanced content.
63
+
64
+ ## How to Get Started with the Model
65
+
66
+ Use the code below to get started with the model:
67
+
68
+ ```python
69
+ from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
70
+
71
+ tokenizer = AutoTokenizer.from_pretrained("path_to_your_model")
72
+ model = AutoModelForSeq2SeqLM.from_pretrained("path_to_your_model")
73
+
74
+ inputs = tokenizer("Translate this text", return_tensors="pt")
75
+ outputs = model.generate(**inputs)
76
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
77
+ ```
78
+
79
+ ## Training Details
80
+
81
+ ### Training Data
82
+
83
+ The model was fine-tuned on a dataset comprising multilingual text pairs for English, Hindi, and Greek. The dataset includes:
84
+ - Publicly available bilingual corpora.
85
+ - Synthetic data for low-resource language pairs.
86
+
87
+ ### Training Procedure
88
+
89
+ #### Preprocessing
90
+
91
+ - Tokenization with `facebook/m2m100_418M` tokenizer.
92
+ - Dynamic padding with sequences padded to the longest in the batch.
93
+
94
+ #### Training Hyperparameters
95
+
96
+ - **Training regime:** Mixed precision (fp32)
97
+ - **Batch size:** 16
98
+ - **Learning rate:** 2e-5
99
+ - **Number of epochs:** 10
100
+
101
+ #### Speeds, Sizes, Times [optional]
102
+
103
+ - Approximate training time: 1218 minutes on M3 Pro chip.
104
+
105
+ ## Evaluation
106
+
107
+
108
+ #### Testing Data
109
+
110
+ The model was evaluated using a held-out test set from the same multilingual dataset used for fine-tuning.
111
+
112
+ #### Metrics
113
+
114
+ The primary evaluation metric is BLEU score.
115
+
116
+ ### Results
117
+
118
+ The model achieved the following BLEU scores:
119
+ - English to Hindi: 36.2
120
+ - English to Greek: 31.5
121
+
122
+ ## Environmental Impact
123
+
124
+ - **Hardware Type:** M3 Pro Chip (MacBook)
125
+ - **Hours used:** ~1218 Min
126
+ - **Compute Region:** Local
127
+ - **Carbon Emitted:** Negligible (powered by renewable energy sources, where applicable)
128
+
129
+ ## Citation [optional]
130
+
131
+ **BibTeX:**
132
+ ```bibtex
133
+ @misc{arun_translation_model,
134
+ author = {Arun Kumar Roy},
135
+ title = {Multilingual Translation Model for English, Hindi, and Greek},
136
+ year = {2025},
137
+ publisher = {Hugging Face}
138
+ }
139
+ ```
140
+
141
+ ## Model Card Authors
142
+
143
+ - Arun Kumar Roy
144
+
145
+ ## Model Card Contact
146
+
147
+ For inquiries, contact [https://github.com/aktheroy].
148
+