Spaces:
Runtime error
Runtime error
mkumar87AI
commited on
Commit
β’
f4d8de3
1
Parent(s):
7c629c6
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,125 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""SimpleChatBot_OpenSourceModel.ipynb
|
3 |
+
|
4 |
+
Automatically generated by Colaboratory.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/1q7EXhcR6gncrcwySFbN7u9fOIwTc4LtD
|
8 |
+
|
9 |
+
##*** Note : *** You will be NOT be charged for this exercise. Everything is OpenSource
|
10 |
+
|
11 |
+
### This notebook presents how to make a simple conversational chatbot using Open Source language model that we will download from hugging-face hub
|
12 |
+
|
13 |
+
### Fix the UTF-8 encoding
|
14 |
+
"""
|
15 |
+
|
16 |
+
import locale
|
17 |
+
locale.getpreferredencoding = lambda: "UTF-8"
|
18 |
+
|
19 |
+
"""### Install the python packages. They are need to execute necessary to make the program work"""
|
20 |
+
|
21 |
+
pip install -qq -U langchain transformers sentence-transformers bitsandbytes accelerate
|
22 |
+
|
23 |
+
!pip install git+https://github.com/huggingface/transformers@v4.31-release
|
24 |
+
|
25 |
+
!pip install transformers -U
|
26 |
+
|
27 |
+
"""### Import the necessary libraries"""
|
28 |
+
|
29 |
+
from langchain.llms.huggingface_pipeline import HuggingFacePipeline
|
30 |
+
import torch
|
31 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
32 |
+
from transformers import BitsAndBytesConfig
|
33 |
+
from langchain.chains import LLMChain
|
34 |
+
from langchain.prompts import PromptTemplate
|
35 |
+
from langchain.embeddings import (OpenAIEmbeddings, HuggingFaceEmbeddings)
|
36 |
+
from langchain.schema import StrOutputParser
|
37 |
+
from langchain.schema.runnable import RunnablePassthrough
|
38 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
39 |
+
from langchain.vectorstores import Chroma
|
40 |
+
from langchain.document_loaders import PyPDFLoader
|
41 |
+
from langchain.callbacks.manager import CallbackManager
|
42 |
+
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
|
43 |
+
|
44 |
+
"""### Select and download the model from Hugging face
|
45 |
+
|
46 |
+
#### Hugging face hub contains a lot of pre-trained AI models related to computer vision, NLP, etc. For our task, we would need to use a text-generation model. Follow the steps below to choose and download a model
|
47 |
+
|
48 |
+
1. Go to this link -> https://huggingface.co/models?pipeline_tag=text-generation&sort=trending
|
49 |
+
|
50 |
+
2. For this example, we will be using the Mitsral 7B Instruct v0.2 [https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2]
|
51 |
+
"""
|
52 |
+
|
53 |
+
torch.set_default_device("cuda")
|
54 |
+
|
55 |
+
model_id = "mistralai/Mistral-7B-Instruct-v0.2"
|
56 |
+
model = AutoModelForCausalLM.from_pretrained(model_id,
|
57 |
+
device_map='auto',
|
58 |
+
torch_dtype="auto",
|
59 |
+
load_in_4bit=True,
|
60 |
+
trust_remote_code=True,
|
61 |
+
low_cpu_mem_usage=True)
|
62 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
|
63 |
+
|
64 |
+
"""### Setup the pipeline using tokenizer and model
|
65 |
+
|
66 |
+
"""
|
67 |
+
|
68 |
+
from transformers import pipeline
|
69 |
+
|
70 |
+
pipe = pipeline(
|
71 |
+
task = "text-generation",
|
72 |
+
model = model,
|
73 |
+
tokenizer = tokenizer,
|
74 |
+
pad_token_id = tokenizer.eos_token_id,
|
75 |
+
temperature = 0.3,
|
76 |
+
top_k = 50,
|
77 |
+
top_p = 0.95,
|
78 |
+
max_new_tokens=3072,
|
79 |
+
repetition_penalty = 1.2
|
80 |
+
)
|
81 |
+
|
82 |
+
"""### Create an llm object"""
|
83 |
+
|
84 |
+
llm = HuggingFacePipeline(pipeline = pipe)
|
85 |
+
|
86 |
+
"""### Create a simple prompt tempelate using Langchain framework"""
|
87 |
+
|
88 |
+
template = """
|
89 |
+
"<s>[INST] You are a question and answering bot
|
90 |
+
You always respond with a funny twist, and keep your
|
91 |
+
answers short. Now answer this Question : {question}.
|
92 |
+
To keep you more stateful, you also get help with previous
|
93 |
+
chat history : {chat_history}[/INST]
|
94 |
+
"""
|
95 |
+
prompt = PromptTemplate(template=template, input_variables=["question", "chat_history"])
|
96 |
+
|
97 |
+
"""### Create an llm chain"""
|
98 |
+
|
99 |
+
llm_chain = LLMChain(prompt=prompt, llm=llm)
|
100 |
+
|
101 |
+
"""### Try invoking the LLM, with a simple chain"""
|
102 |
+
|
103 |
+
def ask_me_chat_completions(query, chat_history, llm_chain):
|
104 |
+
response = llm_chain.run({"question":query,"chat_history":chat_history})
|
105 |
+
return response
|
106 |
+
|
107 |
+
"""### Question and Answer segment
|
108 |
+
|
109 |
+
#### **Activity : ** Try the following things
|
110 |
+
|
111 |
+
1. Search on the internet regarding the context length of the model
|
112 |
+
Check when does the model reach its context limit and stops responding ?
|
113 |
+
2. Can you do anything in the prompt to fit more conversations in the context length ?
|
114 |
+
3. Can you programtically increase the context window ?
|
115 |
+
4. Can you programatically make the memory management better ?
|
116 |
+
"""
|
117 |
+
|
118 |
+
query = None
|
119 |
+
chat_history = []
|
120 |
+
while query != "q":
|
121 |
+
query = input("Ask your questions here, press q to quit: ")
|
122 |
+
if query != "q":
|
123 |
+
response = ask_me_chat_completions(query, chat_history, llm_chain)
|
124 |
+
print(f'Your query returned the following response: {response}')
|
125 |
+
chat_history.append(response)
|