erfanzar commited on
Commit
7c0e3e4
β€’
1 Parent(s): ac2dcee

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +12 -65
README.md CHANGED
@@ -23,84 +23,31 @@ this model uses Task classification and the conversation is between USER and Ans
23
  # NOTE ⚠️
24
 
25
 
26
- THE JAX/FLAX version of model is available both for training and usage
27
 
 
28
 
29
 
30
- # Using Model in Huggingface Transformers
31
 
32
- ## Examples πŸš€
33
-
34
- ```text
35
- </s><|prompter|> TEXT </s><|ai|>
36
  ```
37
 
38
- For Byte by Byte Generation, You can use this code
39
- ```python
40
- # It's recommended to use PipeLine
41
- # Make Sure that you have sentence piece,bits and bytes and accelerate installed
42
- from transformers import LlamaTokenizer, LlamaForCausalLM, pipeline, GenerationConfig
43
- import torch
44
- from IPython.display import clear_output
45
- import textwrap
46
- from typing import List, Optional
47
- import re
48
- import base64
49
-
50
- tokenizer = LlamaTokenizer.from_pretrained("erfanzar/LGeM-7B-MT")
51
- model = LlamaForCausalLM.from_pretrained(
52
- 'erfanzar/LGeM-7B-MT',
53
- load_in_8bit=True,
54
- device_map='auto',
55
- torch_dtype=torch.float16
56
- )
57
-
58
- def generator(input_text,pipe_line,max_number=256,do_print=False ,args_a=False):
59
- verify_text = lambda txt : '\n'.join([textwrap.fill(txt, width=140) for txt in txt.split('\n')])
60
-
61
- orginal_text = input_text
62
- if not input_text.startswith(f'<|prompter|>') and args_a:
63
- input_text = f'<\s><|prompter|> {input_text}<\s><|ai|>'
64
- for i in range(max_number):
65
- exac = input_text
66
- with torch.no_grad():
67
- output = pipe_line(input_text)
68
- input_text = output[0]['generated_text']
69
- if do_print:
70
- clear_output(wait=True)
71
- print(verify_text(input_text))
72
-
73
- if input_text.endswith('<\s>') and i>6 or exac == input_text or input_text.endswith('<|prompter|>') and i>6:
74
- break
75
- yield verify_text(input_text)
76
 
 
 
77
  ```
78
 
79
-
80
- And Use just like
81
-
82
-
83
-
84
- ```python
85
-
86
- pipe_line = pipeline(
87
- "text-generation",
88
- model=model,
89
- tokenizer=tokenizer,
90
- temperature=0.8,
91
- top_p=0.95,
92
- max_new_tokens=4,
93
- output_scores=True
94
-
95
- )
96
- ```
97
  or Just Simply Open [GOOGLE COLAB πŸš€πŸš€](https://colab.research.google.com/drive/1nWS_FhWIDH3-g56F3FbWCIYi0ngVdWHx?usp=sharing)
98
 
99
  ### Generate Method to get res Text by Text
100
 
101
  ```python
102
 
103
- def generate(model_,input_ids_,tokeinzer_,max_length:int=256,temperature :float= 1,eos_token_id:int=2):
104
  with torch.no_grad():
105
  before_start = len(input_ids_[0])+1
106
  for _ in range(max_length):
@@ -160,8 +107,8 @@ from modules import FlaxLGeMForCausalLM
160
 
161
  - and Training code is available at jax_train.py (check source)
162
  - training parameters
163
- - - learning rate 5e-5
164
- - - Optimizer LION
165
  - - batch 32
166
  - - TPU POD
167
  - - Train Time 50 hours
 
23
  # NOTE ⚠️
24
 
25
 
 
26
 
27
+ THE JAX/FLAX version of model is available both for training and usage And This model support context length of 3300
28
 
29
 
30
+ this model support run with OST_UI so heres how to run it with just one command
31
 
32
+ ```shell
33
+ git clone https://github.com/erfanzar/OST-OpenSourceTransformers
34
+ cd OST-OpenSourceTransformers/
35
+ python3 OST_UI/app.py --model_id='erfanzar/chatLGeM' --
36
  ```
37
 
38
+ ## Examples πŸš€
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
 
40
+ ```text
41
+ </s><|prompter|> TEXT </s><|assistant|>
42
  ```
43
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
44
  or Just Simply Open [GOOGLE COLAB πŸš€πŸš€](https://colab.research.google.com/drive/1nWS_FhWIDH3-g56F3FbWCIYi0ngVdWHx?usp=sharing)
45
 
46
  ### Generate Method to get res Text by Text
47
 
48
  ```python
49
 
50
+ def generate(model_,input_ids_,tokeinzer_,max_length:int=3300,temperature :float= 0.2,eos_token_id:int=2):
51
  with torch.no_grad():
52
  before_start = len(input_ids_[0])+1
53
  for _ in range(max_length):
 
107
 
108
  - and Training code is available at jax_train.py (check source)
109
  - training parameters
110
+ - - learning rate 2e-5
111
+ - - Optimizer AdamW
112
  - - batch 32
113
  - - TPU POD
114
  - - Train Time 50 hours