-
Notifications
You must be signed in to change notification settings - Fork 1
/
prp_model.py
131 lines (99 loc) · 5.96 KB
/
prp_model.py
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
import os
import re
import numpy as np
import pandas as pd
import torch
import json
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, TrainingArguments
from peft import LoraConfig
from trl import DPOTrainer
from datasets import Dataset
from tqdm import tqdm
from score import score_relevance, score_APC
def load_generator(prp_scale):
model_id = f"google/gemma-1.1-{prp_scale}-it"
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16
)
prp_tokenizer = AutoTokenizer.from_pretrained(model_id, token=os.environ['HF_TOKEN'])
prp_model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config, device_map={"":0}, token=os.environ['HF_TOKEN'])
return prp_tokenizer, prp_model
def retrieval_augmented_generate(character, statements, query, prp_model, prp_tokenizer, discriminator, rag_top_k):
system_prompt = f"You are an AI agent role-playing as {character}, you should give a short response to the user's utterance as the character, not as an AI."
scores = [score_relevance(character, statement, query, discriminator)[1].item() for statement in statements]
retrieval_augmented_context = "\n".join([statements[idx] for idx in np.argsort(scores)[::-1][:rag_top_k]])
system_prompt = retrieval_augmented_context+"\n"+system_prompt
input_text = f"<start_of_turn>model\n{system_prompt}<end_of_turn>\n<start_of_turn>user\n{query}<end_of_turn>\n<start_of_turn>model\n"
input_ids = prp_tokenizer(input_text, return_tensors="pt").to("cuda")
outputs = prp_model.sample(**input_ids, max_new_tokens=256, temperature=1.0)
response = prp_tokenizer.decode(outputs[0])
if not response.endswith("<eos>"):
response = response + "<eos>"
response = re.findall("\n<start_of_turn>model\n(.*)?<eos>", response, re.DOTALL)[0] + "<eos>"
return input_text, response
def generate_rag_dpo_dataset(character, prp_model, prp_tokenizer, relevance_discriminator, nli_discriminator, persona_statement_dataset, relevant_query_dataset, max_dpo_data, rag_top_k):
if os.path.exists(f"statement/{character}.rag.dpo.json"):
return json.load(open(f"statement/{character}.rag.dpo.json"))
queries = [query for data in relevant_query_dataset for query in data["queries"]]
queries = np.random.choice(queries, min(len(queries), max_dpo_data), replace=False)
statements = [data["statement"] for data in persona_statement_dataset]
dataset = []
bar = tqdm(queries)
with torch.no_grad():
for query in bar:
input_text, response_1 = retrieval_augmented_generate(character, statements, query, prp_model, prp_tokenizer, relevance_discriminator, rag_top_k)
apc_1 = score_APC(character, statements, query, response_1, relevance_discriminator, nli_discriminator).item()
input_text, response_2 = retrieval_augmented_generate(character, statements, query, prp_model, prp_tokenizer, relevance_discriminator, rag_top_k)
apc_2 = score_APC(character, statements, query, response_2, relevance_discriminator, nli_discriminator).item()
if np.abs(apc_1 - apc_2) > 0.2:
chosen = [response_1, response_2][np.argmax([apc_1, apc_2])]
rejected = [response_1, response_2][np.argmin([apc_1, apc_2])]
dataset.append({"prompt": input_text, "chosen": chosen, "rejected": rejected})
bar.set_description(f"Generating APC-based RAG DPO Dataset... Number of Data: {len(dataset)}")
json.dump(dataset, open(f"statement/{character}.rag.dpo.json", "w"))
return dataset
def retrieval_augmented_generate_zh(character, statements, query, prp_model, prp_tokenizer, discriminator, rag_top_k):
system_prompt = f"你是一个扮演“{character}”的AI, 你应当扮演“{character}”回应用户，而不是作为一个AI。"
scores = [score_relevance(character, statement, query, discriminator)[1].item() for statement in statements]
retrieval_augmented_context = "\n".join([statements[idx] for idx in np.argsort(scores)[::-1][:rag_top_k]])
system_prompt = retrieval_augmented_context+"\n"+system_prompt
input_text = f"<start_of_turn>model\n{system_prompt}<end_of_turn>\n<start_of_turn>user\n{query}<end_of_turn>\n<start_of_turn>model\n"
input_ids = prp_tokenizer(input_text, return_tensors="pt").to("cuda")
outputs = prp_model.sample(**input_ids, max_new_tokens=256, temperature=1.0)
response = prp_tokenizer.decode(outputs[0])
if not response.endswith("<eos>"):
response = response + "<eos>"
response = re.findall("\n<start_of_turn>model\n(.*)?<eos>", response, re.DOTALL)[0] + "<eos>"
return input_text, response
def train_prp(character, prp_model, prp_tokenizer, prp_scale, rag_dpo_dataset, lora_rank, prp_dpo_epoch):
lora_config = LoraConfig(
r=lora_rank,
target_modules=["q_proj", "o_proj", "k_proj", "v_proj", "gate_proj", "up_proj", "down_proj"],
task_type="CAUSAL_LM",
)
prp_model.add_adapter(lora_config)
dpo_trainer = DPOTrainer(
prp_model,
ref_model=None,
beta=0.1,
train_dataset=Dataset.from_pandas(pd.DataFrame(rag_dpo_dataset)),
tokenizer=prp_tokenizer,
args=TrainingArguments(
per_device_train_batch_size=1,
gradient_accumulation_steps=8,
num_train_epochs=prp_dpo_epoch,
learning_rate=2e-4,
save_steps=10,
fp16=True,
logging_steps=1,
output_dir=f"prp_models/gemma-1.1-{prp_scale}-it-lora-{character}-rag-dpo",
optim="paged_adamw_8bit"
),
peft_config=lora_config,
)
dpo_trainer.train()
prp_tokenizer.save_pretrained(f"prp_models/gemma-1.1-{prp_scale}-it-lora-{character}-rag-dpo")
prp_model.save_pretrained(f"prp_models/gemma-1.1-{prp_scale}-it-lora-{character}-rag-dpo")
return prp_tokenizer, prp_model