antitheft159
commited on
Commit
•
478ef55
1
Parent(s):
02e8704
Upload laurelstring_gpt2_tttg_159.py
Browse files- laurelstring_gpt2_tttg_159.py +207 -0
laurelstring_gpt2_tttg_159.py
ADDED
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""laurelString/gpt2/tttg.159
|
3 |
+
|
4 |
+
Automatically generated by Colab.
|
5 |
+
|
6 |
+
Original file is located at
|
7 |
+
https://colab.research.google.com/drive/16-bSqq2kMNO8X0BjNA0-bCckjnx1Ler_
|
8 |
+
"""
|
9 |
+
|
10 |
+
! pip install sentence_transformers==2.2.2
|
11 |
+
|
12 |
+
!pip install -qq -U langchain
|
13 |
+
!pip install -qq -U langchaing-community
|
14 |
+
!pip install -qq -U tiktoken
|
15 |
+
!pip install -qq -U pypdf
|
16 |
+
!pip install -qq -U faiss-gpu
|
17 |
+
!pip install -qq -U InstructorEmbedding
|
18 |
+
!pip install -qq -U accelerate
|
19 |
+
!pip install -qq -U bitsandbytes
|
20 |
+
|
21 |
+
import warnings
|
22 |
+
warnings.filterwarnings("ignore")
|
23 |
+
|
24 |
+
import os
|
25 |
+
import glob
|
26 |
+
import textwrap
|
27 |
+
import time
|
28 |
+
|
29 |
+
import langchain
|
30 |
+
|
31 |
+
from langchain.document_loaders import PyPDFLoader, DirectoryLoader
|
32 |
+
|
33 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
34 |
+
|
35 |
+
from langchain import PropmtTemplate, LLMChain
|
36 |
+
|
37 |
+
from langchain.vectorstores import FAISS
|
38 |
+
|
39 |
+
from langchain.llms import HuggingFacePipeline
|
40 |
+
from langchain.embeddings import HuggingFaceInstructEmbeddings
|
41 |
+
|
42 |
+
from langchain.chains import Retrieva1QA
|
43 |
+
|
44 |
+
import torch
|
45 |
+
import transformers
|
46 |
+
from transformers import (
|
47 |
+
AutoTokenizer, AutoModelForCausalLM,
|
48 |
+
BitsAndBytesConfig,
|
49 |
+
pipeline
|
50 |
+
)
|
51 |
+
|
52 |
+
class RAG:
|
53 |
+
temperature = 0,
|
54 |
+
top_p = 0.95,
|
55 |
+
repetition_penalty = 1.15
|
56 |
+
|
57 |
+
split_chunk_size = 800
|
58 |
+
split_overlap = 0
|
59 |
+
|
60 |
+
embeddings_model_repo = 'sentence-transformers/all-MiniLM-L6-v2'
|
61 |
+
|
62 |
+
k = 5
|
63 |
+
|
64 |
+
PDFs_path = '/kaggle/input/physics9thclass/'
|
65 |
+
Embeddings_path = '/kaggle/working/embeddingfinal/'
|
66 |
+
Persist_directory = './books-vectorb'
|
67 |
+
|
68 |
+
model_repo = 'darl149/llama-2-13b-chat-hf'
|
69 |
+
tokenizer = AutoTokenizer.from_pretrained(model_repo, use_fast=True)
|
70 |
+
model = AutoModelForCausalLM.from_pretrained(
|
71 |
+
model_repo,
|
72 |
+
load_in_4bit = True,
|
73 |
+
device_map = 'auto',
|
74 |
+
torch_dtype = torch.float16,
|
75 |
+
low_cpu_mem_usage = True,
|
76 |
+
trust_remote_code = True
|
77 |
+
)
|
78 |
+
|
79 |
+
max_len = 2048
|
80 |
+
|
81 |
+
pipe = pipeline(
|
82 |
+
task = "text-generation",
|
83 |
+
model = model,
|
84 |
+
tokenizer = tokenizer,
|
85 |
+
pad_token_id = tokenizer.eos_token_id,
|
86 |
+
max_length = max_len,
|
87 |
+
temperature = RAG.temperature,
|
88 |
+
top_p = RAG.top_p
|
89 |
+
repetition_penalty = RAG.repetition_penalty
|
90 |
+
)
|
91 |
+
|
92 |
+
llm = HuggingFacePipeline(pipeline = pipe)
|
93 |
+
|
94 |
+
query = """Give me the detail on momentum and torque and how they are different."""
|
95 |
+
llm.invoke(query, truncation=True)
|
96 |
+
|
97 |
+
loader = DircetoryLoader(
|
98 |
+
RAG.Embeddings_path,
|
99 |
+
glob="./*.pdf",
|
100 |
+
loader_cls=PyPDFLoader,
|
101 |
+
show_progress=True,
|
102 |
+
use_multithreading=True
|
103 |
+
)
|
104 |
+
|
105 |
+
documents = loader.load()
|
106 |
+
|
107 |
+
print(f'We have {len(documents)} pages in total')
|
108 |
+
|
109 |
+
documents[100].page_content
|
110 |
+
|
111 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
112 |
+
chunk_size = RAG.split_chunk_size,
|
113 |
+
chunk_overlap = RAG.split_documents(documents)
|
114 |
+
|
115 |
+
print(f'We have created {len(texts)} chunks from {len(documents)} pages')
|
116 |
+
)
|
117 |
+
|
118 |
+
if not os.path.exists(RAG.Embeddings_path + '/index.faiss'):
|
119 |
+
|
120 |
+
embeddings = HuggingFaceInstructEmbeddings(
|
121 |
+
model_name = RAG.embeddings_model_repo,
|
122 |
+
model_kwargs = {"device": "cuda"}
|
123 |
+
)
|
124 |
+
vectordb.save_local(f"{RAG.Persist_directory}/faiss_index_hp")
|
125 |
+
|
126 |
+
embeddings = HuggingFaceInstructEmbeddings(
|
127 |
+
model_name = RAG.embeddings_model_repo,
|
128 |
+
model_kwargs = {"device": "cuda"}
|
129 |
+
)
|
130 |
+
|
131 |
+
vectordb = FAISS.load_local(
|
132 |
+
RAG.Persist_directory + '/faiss_index_hp',
|
133 |
+
embeddings,
|
134 |
+
allow_dangerous_deserialization=True
|
135 |
+
)
|
136 |
+
|
137 |
+
vectordb.similarity_search('quantum')
|
138 |
+
|
139 |
+
prompt_template = """Suppose you are a Teaching assitant.
|
140 |
+
Your task is to gave answers to the asked questions with sympathy, empathy and kind words.
|
141 |
+
Start by something like good question or very good point etc.
|
142 |
+
Ensure your response is directed at the person asking the question, assuming they are not another teacher but a student seeking guidance.
|
143 |
+
At the end of the answer, give best wishe like "I hope you understand. If not, I'll be glad to explain to you again,"
|
144 |
+
Please try to be as concise as you can and use no more words than 150.
|
145 |
+
Important Note: Please provide as accurate answers as you can and for numerical problems provide explanation.
|
146 |
+
Try to follow the following pieces of context as much as you can but you can also use your own information.
|
147 |
+
|
148 |
+
{context}
|
149 |
+
|
150 |
+
Question: {question}
|
151 |
+
Answer:"""
|
152 |
+
|
153 |
+
PROMPT = PrompTemplate(
|
154 |
+
template = prompt_template,
|
155 |
+
input_variables = ["context", "question"]
|
156 |
+
)
|
157 |
+
|
158 |
+
retriver = vectordb.as_retriever(search_kwargs = {
|
159 |
+
"k": RAG.k, "search_type" : "similarity"})
|
160 |
+
|
161 |
+
qa_chain = RetrievalQA.from_chain_type(
|
162 |
+
llm = llm,
|
163 |
+
chain_type = "stuff", # map_reduce, map_rerank,stuff, refine
|
164 |
+
retriever = retriever,
|
165 |
+
chain_type_kwargs = {"prompt": PROMPT},
|
166 |
+
return_source_documents = True,
|
167 |
+
verbose = False
|
168 |
+
)
|
169 |
+
|
170 |
+
question = "First law of motion has another name what it is."
|
171 |
+
vectordb.max_marginal_relevance_search(question, k = RAG.k)
|
172 |
+
|
173 |
+
def wrap_text_preserve_newlines(text, width=700):
|
174 |
+
lines = text.split('\n')
|
175 |
+
|
176 |
+
wrapped_lines = [textwrap.fill(line, width=width) for line in lines]
|
177 |
+
|
178 |
+
wrapped_text = '\n'.join(wrapped_lines)
|
179 |
+
|
180 |
+
return wrapped_text
|
181 |
+
|
182 |
+
def process_llm_response(llm_response):
|
183 |
+
answer_full = llm_response['result']
|
184 |
+
answer_start = answer_full.find("Answer:") + 1en("Answer:")
|
185 |
+
answer = answer_full[answer_start:].strip()
|
186 |
+
|
187 |
+
answer = wrap_text_preserve_newlines(answer)
|
188 |
+
return answer
|
189 |
+
|
190 |
+
def llm_ans(query):
|
191 |
+
|
192 |
+
llm_response = qa_chain.invoke(query)
|
193 |
+
ans = process_llm_response(llm_response)
|
194 |
+
end = time.time()
|
195 |
+
|
196 |
+
return ans
|
197 |
+
|
198 |
+
query = "Firt law of motion has another name what it is."
|
199 |
+
print(llm_ans(query))
|
200 |
+
|
201 |
+
query = """Firt law of motion has another name what it is."""
|
202 |
+
llm.invoke(query,truncation=True)
|
203 |
+
|
204 |
+
query = "The concrete roof of a house of thickness 20 cm has an area 200 m2. The temperature inside the house is 15° C and outside is 35° C. find the rate at which thermal energy conducted through the roof in Js-1. The value of k for concrete is 0.65 Wm1K1."
|
205 |
+
print(llm_ans(query))
|
206 |
+
|
207 |
+
query = """The concrete roof of a house of thickness 20 cm has an area 200 m2. The temperature inside the house is 15° C and outside is 35° C. find the rate at which thermal energy conducted through the roof in Js-1. The value of k for concrete is 0.65 Wm1K1."""
|