-
Notifications
You must be signed in to change notification settings - Fork 0
/
RLC.py
259 lines (225 loc) · 7.5 KB
/
RLC.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
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
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
import os
from typing import List, Dict
import argparse
from datasets import load_dataset
from tqdm import tqdm
from transformers import AutoTokenizer
from accelerate import Accelerator
import trlx
from trlx.data.configs import (
ModelConfig,
OptimizerConfig,
SchedulerConfig,
TokenizerConfig,
TrainConfig,
TRLConfig,
)
from trlx.models.modeling_ppo import PPOConfig
import trlx.utils.logging as logging
from prompts import CHAIN_OF_THOUGHT_PROMPTS, OPENAI_API_KEY
from chatbot.chatbot import BaseBot, ChatGPTBot, T5, BART, GPT2
from utils.utils import *
try:
import evaluate
except ImportError:
raise ImportError(
"To run this example, please install the `evaluate` and `nltk` packages" "by running `pip install evaluate`"
)
def save(path):
if not os.path.exists(path):
os.makedirs(path)
trainer.save_pretrained(path)
meteor = evaluate.load("meteor")
if __name__ == "__main__":
accelerator = Accelerator()
parser = argparse.ArgumentParser()
parser.add_argument(
"--model_name",
type=str,
default='google/flan-t5-xl',
help="which model to load",
)
parser.add_argument(
"--ask_mode",
type=str,
help="how to ask for judgement",
default='stadnard_answer_reward'
)
parser.add_argument(
"--is_chain_of_thought",
type=bool,
help="ask with chain of thought or not for judgement",
default=False
)
parser.add_argument(
"--bbh_set",
type=str,
help="BBH set name",
default="date_understanding"
)
parser.add_argument(
"--seed",
type=int,
default=1000,
)
parser.add_argument(
"--cache_dir",
type=str,
default=".cache/",
)
parser.add_argument(
"--answer_num",
type=int,
default=1,
)
args = parser.parse_args()
config = TRLConfig(
train=TrainConfig(
seq_length=512,
epochs=100,
total_steps=8000,
batch_size=12,
checkpoint_interval=1000,
eval_interval=50,
pipeline="PromptPipeline",
trainer="AcceleratePPOTrainer",
seed = args.seed
),
model=ModelConfig(
model_path=args.model_name,
model_arch_type="seq2seq",
num_layers_unfrozen=2,
),
tokenizer=TokenizerConfig(
tokenizer_path=args.model_name,
truncation_side="right",
),
optimizer=OptimizerConfig(
name="adamw",
kwargs={
"lr": 1.0e-4,
"betas": [0.9, 0.999],
"eps": 1.0e-8,
"weight_decay": 1.0e-6,
},
),
scheduler=SchedulerConfig(
name="cosine_annealing",
kwargs={
"T_max": 10000,
"eta_min": 1.0e-6,
},
),
method=PPOConfig(
name="PPOConfig",
num_rollouts=512,
chunk_size=12,
ppo_epochs=4,
init_kl_coef=0.1,
target=6,
horizon=10000,
gamma=0.99,
lam=0.95,
cliprange=0.2,
cliprange_value=0.2,
vf_coef=1.0,
scale_reward=None,
ref_mean=None,
ref_std=None,
cliprange_reward=10,
gen_kwargs={
"max_new_tokens": 100,
},
gen_experience_kwargs={
"max_new_tokens": 100,
"do_sample": True,
"temperature": 1.0,
"top_k": 50,
"top_p": 0.95,
},
),
)
logger = logging.get_logger()
logger.info(f"BBH set name: {args.bbh_set}")
MODELS: Dict[str, BaseBot] = {
"google/flan-t5-small": T5,
"google/flan-t5-base": T5,
"google/flan-t5-large": T5,
"google/flan-t5-xl": T5,
"google/flan-t5-xxl": T5,
"facebook/bart-large": BART,
"facebook/bart-base": BART,
"gpt2-large": GPT2,
"gpt2-medium": GPT2,
"chatgpt": ChatGPTBot,
}
judge_model = MODELS[args.model_name](args) if args.model_name != "chatgpt" else MODELS[args.model_name](OPENAI_API_KEY)
def reward_fn(samples: List[str], prompts: List[str], outputs: List[str]):
targets = [prompt_label[prompt.strip()].lower() for prompt in prompts]
scores = []
for index in range(len(prompts)):
assert len(prompts) == len(outputs), f"Different length :target:{targets}, outputs:{outputs}"
question = prompts[index]
question = question.split("Tell me the options directly, excluding the content of the options")[0]
if args.ask_mode == 'standard_answer_reward':
answer = outputs[index].lower()
if targets[index] in answer:
scores.append(1)
else:
scores.append(0)
elif args.ask_mode == 'judge-direct':
answer = outputs[index]
prompt = get_judge_prompt(args.ask_mode, question, answer)
judge_score = judge_model.ask(prompt)
if len(judge_score) == 0:
print(f"judge_score is empty: {judge_score}")
scores.append(0)
continue
judge_answer = judge_score.lower()
if 'yes' in judge_answer:
scores.append(1)
else:
scores.append(0)
return scores
def metric_fn(samples: List[str], prompts: List[str], outputs: List[str]) -> List[float]:
"""Compute COMET, BLEU and CHRF for evaluation"""
targets = [prompt_label[prompt.strip()].lower() for prompt in prompts]
scores = 0.0
for index in range(len(prompts)):
answer = outputs[index].lower()
if targets[index] in answer:
scores += 1.0
return {"score": scores/len(prompts)}
dataset = load_dataset("lukaemon/bbh", args.bbh_set)
prompts = dataset["test"]["input"]
targets = dataset["test"]["target"]
if args.is_chain_of_thought:
prompts = [prompt+CHAIN_OF_THOUGHT_PROMPTS for prompt in prompts]
else:
prompts = [prompt for prompt in prompts]
tokenizer = AutoTokenizer.from_pretrained(config.model.model_path)
tokenizer.padding_side = "left"
tokenizer.truncation_side = "right"
tokenizer.sep_token = "<sep>"
prompt_label = {}
max_length = config.train.seq_length - config.method.gen_kwargs["max_new_tokens"]
for i in tqdm(range(len(prompts))):
key = tokenizer.decode(
tokenizer(prompts[i], truncation=True, max_length=max_length, add_special_tokens=False)["input_ids"],
skip_special_tokens=True,
)
prompt_label[key.strip()] = targets[i]
length = len(prompts)
train_prompts = prompts[:int(0.8*length)]
save_args = {'ask_mode':args.ask_mode, 'save_dataset_name': args.bbh_set, 'save_model_name': args.model_name,'save_is_chain_of_thought': args.is_chain_of_thought, 'seed':args.seed}
trainer = trlx.train(
reward_fn=reward_fn,
prompts=train_prompts,
metric_fn=metric_fn,
eval_prompts=prompts,
config=config,
save_args=save_args,
)
file_to_save = "-".join([str(_arg) for _arg in save_args.values()])
path = "./results/rl_train/" + file_to_save
save(path)