Spaces:
Sleeping
Sleeping
File size: 5,646 Bytes
227a7da 23aeabb 227a7da a70f9b1 227a7da 93a26b1 227a7da 93a26b1 227a7da 8bfe948 227a7da 859bf4c 93a26b1 227a7da 93a26b1 227a7da 93a26b1 8bfe948 93a26b1 227a7da 93a26b1 227a7da 93a26b1 227a7da f439690 227a7da 9b1fd32 227a7da |
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 |
from langchain_community.vectorstores import Chroma
from langchain_community.embeddings import HuggingFaceBgeEmbeddings
from langchain.docstore.document import Document
from transformers import AutoTokenizer
from config import *
import os
import torch
from predibase import Predibase, FinetuningConfig, DeploymentConfig
pb_auth = os.environ.get("pb_token")
pb = Predibase(api_token='pb_8PhUywZ5VQCQFmGl8ZgfiQ')
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})
lorax_client = pb.deployments.client("llama-3-8b-instruct") # Insert deployment name here
lora_weights_rec = REC_LORA_MODEL
lora_weights_exp = EXP_LORA_MODEL
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)
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=20,
batch_size=1,
**kwargs,
):
resp = lorax_client.generate(prompt,adapter_id='wt3639/Llama-3-8B-Instruct_CourseRec_lora', adapter_source='hub', max_new_tokens=max_new_tokens)
return resp
def compare_docs_with_context(doc_a, doc_b, target_occupation_name, target_occupation_dsp,skill_gap):
#courses = f"First: name: {doc_a.metadata['name']} description:{doc_a.metadata['description']} Second: name: {doc_b.metadata['name']} description:{Sdoc_b.metadata['description']}"
courses = f"First: name: {doc_a.metadata['name']} learning outcomes:{doc_a.metadata['skills'][:1500]} Second: name: {doc_b.metadata['name']} learning outcomes:{doc_b.metadata['skills'][:1500]}"
target_occupation = f"name: {target_occupation_name} description: {target_occupation_dsp[:1500]}"
skill_gap = skill_gap
prompt = generate_prompt(target_occupation, skill_gap, courses)
prompt = prompt
output = evaluate(prompt)
# Compare based on the response: [A] means doc_a > doc_b, [B] means doc_a < doc_b
print(output)
result_token_id = output.details.tokens[0].id
if result_token_id == first_id:
return 1 # doc_a should come before doc_b
elif result_token_id == second_id:
return -1 # doc_a should come after doc_b
else:
return 0 # Consider them equal if the response is unclear
#-----------------------------------------explanation-------------------------------------
def generate_prompt_exp(input_text):
return f"""
### Instruction:
As an education expert, you have been provided with information on target occupations and skills gaps, along with recommended course details. Your task is to explain the recommendation in German, focusing on how the course's learning outcomes and target skills relate to the identified skills gaps.
### Input:
{input_text}
### Response:
"""
def generate_exp(
prompt=None,
temperature=0.2,
top_p=0.5,
top_k=20,
num_beams=1,
max_new_tokens=512,
batch_size=1,
do_sample=True,
**kwargs,
):
resp = lorax_client.generate(prompt,adapter_id=EXP_LORA_MODEL, adapter_source='hub', max_new_tokens=max_new_tokens)
return resp.generated_text
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 |