NeMo
PyTorch
English
Hindi
nemotron
ravirajoshi commited on
Commit
7624c49
1 Parent(s): ac81e1e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +132 -6
README.md CHANGED
@@ -1,6 +1,132 @@
1
- ---
2
- license: other
3
- license_name: nvidia-open-model-license
4
- license_link: >-
5
- https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf
6
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: nvidia-open-model-license
4
+ license_link: >-
5
+ https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf
6
+ language:
7
+ - en
8
+ - hi
9
+ base_model: nvidia/Minitron-4B-Base
10
+ library_name: nemo
11
+ ---
12
+
13
+ # Model Overview
14
+
15
+ Nemotron-4-Mini-Hindi-4B-Base is a base model pre-trained on Hindi and English corpus. The Nemotron-Mini-4B-Base (Minitron-4B) is subject to continuous pre-training using Hindi and English data exclusively to create a strong base model for Hindi. It is a small language model (SLM) optimized through distillation, pruning, and quantization for speed and on-device deployment. VRAM usage has been minimized to approximately 2 GB, providing significantly faster time to first token compared to LLMs.
16
+ Please refer to our [arXiv paper](https://arxiv.org/abs/2410.14815) for more details.
17
+
18
+ This model is for research and development only.
19
+
20
+ **Model Developer:** NVIDIA
21
+
22
+ **Model Dates:** Nemotron-4-Mini-Hindi-4B-Base was trained between June 2024 and Oct 2024.
23
+
24
+ ## License
25
+
26
+ Nemotron-4-Mini-Hindi-4B-Base is released under the [NVIDIA Open Model License Agreement](https://developer.download.nvidia.com/licenses/nvidia-open-model-license-agreement-june-2024.pdf).
27
+
28
+ ## Model Architecture
29
+
30
+ Nemotron-4-Mini-Hindi-4B-Base uses a model embedding size of 3072, 32 attention heads, and an MLP intermediate dimension of 9216.
31
+ It also uses Grouped-Query Attention (GQA) and Rotary Position Embeddings (RoPE).
32
+
33
+ **Architecture Type:** Transformer Decoder (auto-regressive language model)
34
+
35
+ **Network Architecture:** Nemotron-4
36
+
37
+ **Input Type:** Text
38
+
39
+ **Input Format:** String
40
+
41
+ **Input Parameters:** None
42
+
43
+ **Other Properties Related to Input:** None
44
+
45
+ **Output Type:** Text
46
+
47
+ **Output Format:** String
48
+
49
+ **Output Parameters:** None
50
+
51
+ **Other Properties Related to Output:** None
52
+
53
+ ## Usage
54
+
55
+ Support for Nemotron models is added to the [transformers library](https://huggingface.co/docs/transformers/en/model_doc/nemotron)
56
+
57
+
58
+ The following code provides an example of how to load the Nemotron-4-Mini-Hindi-4B-Base model and use it to perform text generation.
59
+
60
+ ```python
61
+ import torch
62
+ from transformers import AutoTokenizer, AutoModelForCausalLM
63
+
64
+ # Load the tokenizer and model
65
+ model_path = 'nvidia/Nemotron-4-Mini-Hindi-4B-Base'
66
+ tokenizer = AutoTokenizer.from_pretrained(model_path)
67
+
68
+ device = 'cuda'
69
+ dtype = torch.bfloat16
70
+ model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=dtype, device_map=device)
71
+
72
+ # Prepare the input text
73
+ prompt = 'भारत की संस्कृति के बारे में बताएं।'
74
+ inputs = tokenizer.encode(prompt, return_tensors='pt').to(model.device)
75
+
76
+ # Generate the output
77
+ outputs = model.generate(inputs, max_length=20)
78
+
79
+ # Decode and print the output
80
+ output_text = tokenizer.decode(outputs[0])
81
+ print(output_text)
82
+ ```
83
+
84
+ ## Dataset & Training
85
+
86
+ **Data Collection Method:** Hybrid
87
+
88
+ **Labeling Method:** Not Applicable
89
+
90
+ **Properties:** The Nemotron-Mini-4B base model is used for continuous pre-training. The Nemotron-Mini-4B model is derived from the Nemotron-15B model using compression techniques such as pruning and distillation, consisting of 2.6B trainable parameters. Re-training is performed using a standard causal modeling objective. The dataset consists of 400B tokens, with an equal mix of Hindi and English. The corpus spans domains including legal, math, science, finance, and more. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials.
91
+
92
+ **Data Freshness:** The pre-training data has a cutoff of June 2023.
93
+
94
+ ## Evaluation Results
95
+
96
+ *Zero-shot performance.* Evaluated using select datasets from the [IndicInstruct](https://github.com/AI4Bharat/IndicInstruct) with additions:
97
+
98
+ | MMLU | ARC-C | ARC-E | HellaSwag | BoolQ |
99
+ | :------------- | :------------- | :------------- | :------------- | :------------- |
100
+ | 49.89 | 65.02 | 78.81 | 31.66 | 71.71 |
101
+
102
+
103
+ Please refer to our [paper](https://arxiv.org/abs/2410.14815) for the full set of results.
104
+
105
+ ## Inference
106
+ **Engine:** TensorRT-LLM
107
+
108
+ **Test Hardware:** NVIDIA A100
109
+
110
+ **DType:** Float16/BFloat16
111
+
112
+ ## Limitations
113
+
114
+ The model was trained on data that contains toxic language and societal biases, originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses, especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive. The model may answer with I statements, exhibiting some anthropomorphizing. This issue could be exacerbated without the use of the recommended prompt template.
115
+
116
+ ## Ethical Considerations
117
+
118
+ NVIDIA believes Trustworthy AI is a shared responsibility, and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
119
+
120
+
121
+ ## Citation
122
+
123
+ If you find our work helpful, please consider citing our paper:
124
+ ```
125
+ @article{hindiminitron2024,
126
+ title={Adapting Multilingual LLMs to Low-Resource Languages using Continued Pre-training and Synthetic Corpus},
127
+ author={Raviraj Joshi and Kanishk Singla and Anusha Kamath and Raunak Kalani and Rakesh Paul and Utkarsh Vaidya and Sanjay Singh Chauhan and Niranjan Wartikar and Eileen Long},
128
+ journal={arXiv preprint arXiv:2410.14815},
129
+ year={2024},
130
+ url={https://arxiv.org/abs/2410.14815},
131
+ }
132
+ ```