Spaces:
Paused
Paused
File size: 6,469 Bytes
3430cbb 6b463ef 3430cbb 6b463ef 3430cbb 6b463ef 3430cbb 6b463ef babde94 6b463ef 50220da 6b463ef 3310e88 f6aa47a 6b463ef f6aa47a 6b463ef f6aa47a 6b463ef 3430cbb |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 |
from langchain_community.vectorstores import Chroma
from langchain_community.embeddings import HuggingFaceBgeEmbeddings
from langchain.docstore.document import Document
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
from peft import PeftModel
from config import *
import os
import torch
if torch.cuda.is_available():
device = "cuda"
else:
device = "cpu"
os.environ['CURL_CA_BUNDLE'] = ""
embedding_int = HuggingFaceBgeEmbeddings(
model_name=MODEL_NAME,
encode_kwargs=ENCODE_KWARGS,
query_instruction=QUERY_INSTRUCTION
)
embedding_sim = HuggingFaceBgeEmbeddings(
model_name=MODEL_NAME,
encode_kwargs=ENCODE_KWARGS,
query_instruction='Retrieve semantically similar text.'
)
db = Chroma(persist_directory=PERSIST_DIRECTORY, embedding_function=embedding_int)
retriever = db.as_retriever(search_kwargs={"k": TOP_K})
LLM_MODEL = "meta-llama/Meta-Llama-3-8B-Instruct"
lora_weights = "wt3639/Llama-3-8B-Instruct_CourseRec_lora"
hf_auth = os.environ.get("hf_token")
tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL, token=hf_auth)
first_token = 'First'
second_token = 'Second'
# 获取token的ID
first_id = tokenizer.convert_tokens_to_ids(first_token)
second_id = tokenizer.convert_tokens_to_ids(second_token)
model = AutoModelForCausalLM.from_pretrained(
LLM_MODEL,
torch_dtype=torch.float16,
device_map="auto",
token=hf_auth,
)
rec_adapter = PeftModel.from_pretrained(
model,
lora_weights,
torch_dtype=torch.float16,
device_map={'': 0}
)
tokenizer.padding_side = "left"
# unwind broken decapoda-research config
#model.half() # seems to fix bugs for some users.
rec_adapter.eval()
rec_adapter.config.pad_token_id = tokenizer.pad_token_id = 0 # unk
rec_adapter.config.bos_token_id = 1
rec_adapter.config.eos_token_id = 2
def generate_prompt(target_occupation, skill_gap, courses):
return f"""
### Instruction:
"As an education expert, you have been provided with a target occupation, a skill gap, and information on two candidate courses. Your task is to determine which course better matches the target occupation and skill gap. Please respond with 'First' or 'Second' to indicate your recommendation.
### Input:
Target Occupation: {target_occupation}
Skill Gap: {skill_gap}
candidate courses: {courses}
### Response:
"""
'''
prompt_re = ChatPromptTemplate.from_template(template_re)
chain_re = (
runnable
| prompt_re
)
'''
def evaluate(
prompt=None,
temperature=0,
top_p=1.0,
top_k=40,
num_beams=1,
max_new_tokens=120,
batch_size=1,
**kwargs,
):
inputs = tokenizer(prompt, return_tensors="pt", padding=True, truncation=True).to(device)
generation_config = GenerationConfig(
temperature=temperature,
top_p=top_p,
top_k=top_k,
num_beams=num_beams,
**kwargs,
)
with torch.no_grad():
generation_output = model.generate(
**inputs,
generation_config=generation_config,
return_dict_in_generate=True,
output_scores=True,
max_new_tokens=max_new_tokens,
# batch_size=batch_size,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.eos_token_id,
)
scores = generation_output.scores[0].softmax(dim=-1)
logits = torch.tensor(scores[:,[first_id, second_id]], dtype=torch.float32).softmax(dim=-1)
s = generation_output.sequences
output = tokenizer.batch_decode(s, skip_special_tokens=True)
output = [_.split('Response:\n')[-1] for _ in output]
return output, logits.tolist()
def compare_docs_with_context(doc_a, doc_b, df_course, target_occupation_name, target_occupation_dsp,skill_gap):
# Extract course details from the data frame
course_a = df_course[df_course['course_id'] == int(doc_a.metadata['id'])].iloc[0]
course_b = df_course[df_course['course_id'] == int(doc_b.metadata['id'])].iloc[0]
print('comapring...')
print(course_a['course_name'], course_b['course_name'])
# Prepare the input for chain_re.invoke
courses = f"First: name: {course_a['course_name']} description:{course_a['course_content_limited']} Second: name: {course_b['course_name']} description:{course_b['course_content_limited']}"
#courses = f"First: name: {course_a['course_name']} skills:{course_a['course_skills_edu']} Second: name: {course_b['course_name']} skills:{course_b['course_skills_edu']}"
target_occupation = f"name: {target_occupation_name} description: {target_occupation_dsp}"
skill_gap = skill_gap
prompt = generate_prompt(target_occupation, skill_gap, courses)
prompt = [prompt]
output, logit = evaluate(prompt)
# Compare based on the response: [A] means doc_a > doc_b, [B] means doc_a < doc_b
print(output, logit)
if logit[0][0] > logit[0][1]:
return 1 # doc_a should come before doc_b
elif logit[0][0] < logit[0][1]:
return -1 # doc_a should come after doc_b
else:
return 0 # Consider them equal if the response is unclear
def find_similar_occupation(target_occupation_query, berufe, top_k, similarity_func):
# Pro kurs wird ein Document erstellt. Dieses enthält Metadaten sowie einen page_content.
# Der Inhalt von page_content wird embedded und so für die sucher verwendet.
docs = []
for index, beruf in berufe.iterrows():
# Create document.
doc = Document(
page_content= beruf['short name'] + ' ' + beruf['full name'] + ' ' + beruf['description'],
metadata={
"id": beruf["id"],
"name": beruf['short name'],
"description": beruf["description"],
"entry_requirements": beruf["entry requirements"]
},
)
docs.append(doc)
db_temp = Chroma.from_documents(documents = docs, embedding= embedding_sim, collection_metadata = {"hnsw:space": similarity_func})
# Retriever will search for the top_5 most similar documents to the query.
retriever_temp = db_temp.as_retriever(search_kwargs={"k": top_k})
top_similar_occupations = retriever_temp.get_relevant_documents(target_occupation_query)
return top_similar_occupations |