File size: 5,862 Bytes
9378dcc f0cffd2 7b38588 5d490d5 207fcd8 f0cffd2 3bfb3d7 f0cffd2 0424a04 f0cffd2 207fcd8 7b4a9b9 f0cffd2 8c13143 f0cffd2 7b38588 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 |
---
library_name: transformers
license: mit
language:
- en
- vi
pipeline_tag: text-generation
base_model: HuggingFaceH4/zephyr-7b-beta
tags:
- ghost
---
# Model Card for Model ID
**Ghost 7B Alpha, flying, v0.7.0**
## Model Details
### Model Description
This model is fine tuned from **HuggingFaceH4/zephyr-7b-beta** on a small synthetic datasets (about 200MB) for 50% English and 50% Vietnamese.
- **Developed by:** **Lam H**
- **Language(s) (NLP):** English, Vietnamese
- **License:** MIT
- **Finetuned from model:** [HuggingFaceH4/zephyr-7b-beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta)
## Uses
This model supports both conversation chat and tasks. Feel free to experiment and don't limit your creativity.
The simplest way to try it is to use the `pipeline` from `transformers`.
```python
import torch
from transformers import pipeline
pipe = pipeline(
"text-generation",
model="lamhieu/ghost-7b-v0.9.0",
torch_dtype=torch.bfloat16,
)
```
You can then try any of the sample codes below, formatted using the chat template.
```python
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "nói tôi biết bệnh dịch hạch ở châu Âu do khuẩn nào gây ra"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
tokenized = pipe.tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
outputs = pipe.model.generate(**tokenized, max_new_tokens=512)
results = tokenizer.batch_decode(outputs)[0]
print(results)
# Bệnh dịch hạch ở châu Âu do khuẩn gây ra là do khuẩn Yersinia pestis.
```
```python
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Thông tin về Peristernia despecta"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
tokenized = pipe.tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
outputs = pipe.model.generate(**tokenized, max_new_tokens=512)
results = tokenizer.batch_decode(outputs)[0]
print(results)
# Peristernia despecta là một loài ốc biển, là động vật thân mềm chân bụng sống ở biển trong họ Fasciolariidae.
# ...
```
```python
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "do u know vietnam ?"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
tokenized = pipe.tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
outputs = pipe.model.generate(**tokenized, max_new_tokens=512)
results = tokenizer.batch_decode(outputs)[0]
print(results)
# Yes, I have knowledge about Vietnam. Vietnam is a country in Southeast Asia, bordered by China to the north, Laos and Cambodia to the west, and the South China Sea to the east and south. Its capital city is Hanoi, and its largest city is Ho Chi Minh City (formerly known as Saigon). Vietnam has a population of approximately 100 million people and a diverse cultural heritage influenced by both Chinese and French colonialism. The country has a rich history, including periods of independence, colonization, and resistance, and has experienced significant economic growth in recent years.
```
```python
messages = [
{"role": "system", "content": "You are a helpful assistant, who always provide explanation. Think like you are answering to a five year old."},
{"role": "user", "content": "Tôi yêu em nhiều hơn em nghĩ.\n\nWhich language is this?"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
tokenized = pipe.tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
outputs = pipe.model.generate(**tokenized, max_new_tokens=512)
results = tokenizer.batch_decode(outputs)[0]
print(results)
# This is Vietnamese language. Vietnamese is a language spoken mainly in Vietnam and by the Vietnamese diaspora in many other countries. The sentence you provided means "I love you more than you think." It's like you have more love for someone than they realize.
```
Another example of what you can use to chat multiple turns.
```python
messages = [
# {"role": "system", "content": "You are a helpful and knowledgeable assistant. You like to help and always give honest information, in its original language. In communication, you are always respectful, equal and promote positive behavior."},
{"role": "system", "content": "You are a helpful assistant."}, # Describe to your assistant, anything.
{"role": "user", "content": "Bla bla bla"},
{"role": "assistant", "content": "Bla bla bla"},
{"role": "user", "content": "Bla bla bla"},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
tokenized = pipe.tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
outputs = pipe.model.generate(**tokenized, max_new_tokens=512)
results = tokenizer.batch_decode(outputs)[0]
print(results)
```
## Evaluation
### Results
#### OpenLLM Leaderboard
It will be evaluated and updated later.
#### VMLU
Below are the results evaluated with the VMLU evaluation suite, which is often used to evaluate models that work with Vietnamese.
Note: the results are run with the model in 4bit quantum form, I'm not sure if it has any loss in results or not, if someone can help me run it with full it would be great.
## More Information
Many thanks for
- Datasets: [5CD-AI](https://huggingface.co/5CD-AI), [vilm](https://huggingface.co/vilm).
- Library: [unsloth](https://github.com/unslothai/unsloth)
## Model Card Contact
**Lam H** ([email protected]) |