forked from qibin0506/Cortex
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtrain_ppo.py
More file actions
296 lines (237 loc) · 11.7 KB
/
Copy pathtrain_ppo.py
File metadata and controls
296 lines (237 loc) · 11.7 KB
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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
import os
import re
import concurrent.futures
from typing import List, Optional
import torch
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
from openai import OpenAI
from llm_trainer import PPOTrainer, TrainerTools
from utils import init_env, get_ppo_config, get_eval_prompt
init_env()
# ==========================================
# LLM Judge 配置 (兼容 OpenAI SDK)
# ==========================================
JUDGE_API_KEY = os.environ.get("JUDGE_API_KEY", "")
JUDGE_BASE_URL = os.environ.get("JUDGE_BASE_URL", "https://api.siliconflow.cn/v1")
JUDGE_MODEL = os.environ.get("JUDGE_MODEL", "deepseek-ai/DeepSeek-V3.2")
assert JUDGE_API_KEY != ''
assert JUDGE_BASE_URL != ''
assert JUDGE_MODEL != ''
client = OpenAI(api_key=JUDGE_API_KEY, base_url=JUDGE_BASE_URL)
def call_llm_judge(prompt: str, answer: str, is_think_required: bool) -> float:
"""
调用 LLM API 进行打分，极简 Prompt 防止裁判废话，包含安全拦截机制
"""
safe_answer = answer.replace("<think>", "[思考开始]") \
.replace("</think>", "[思考结束]") \
.replace("<answer>", "[回答开始]") \
.replace("</answer>", "[回答结束]")
base_prompt = """你是一个冷酷客观的 AI 打分机器。你需要评估一个极小AI模型的输出格式与逻辑。
请【完全忽略】事实准确性！只看格式、排版和语言连贯性。
你可以先用一两句话简评，但【文章最后必须且只能】输出一个严格的 XML 标签：<score>得分</score>。范围是 -5.0 到 5.0。"""
if is_think_required:
rule_prompt = """该题【必须包含思考过程】。标准结构：[思考开始]思考内容[思考结束][回答开始]回答内容[回答结束]。
[4.0 到 5.0]：[思考开始]与[思考结束]之间内容饱满，用了"首先/其次/最后"等词，结构清晰。
[2.0 到 3.9]：格式正确，语言基本连贯。
[0.0 到 1.9]：格式正确，但思考过程太短。
[-2.0 到 -0.1]：有轻微复读或逻辑结巴。
[-4.0 到 -2.1]：存在严重的复读废话、车轱辘话。
[-5.0]：乱码，缺少任何一个标识符，或完全答非所问。"""
else:
rule_prompt = """该题【绝对不能思考】！标准结构必须是：[思考开始][思考结束][回答开始]回答内容[回答结束]。
[4.0 到 5.0]：[思考开始]与[思考结束]之间完全为空！回答排版精美。
[2.0 到 3.9]：[思考开始]与[思考结束]之间完全为空！语言连贯。
[0.0 到 1.9]：[思考开始]与[思考结束]之间完全为空！但回答简略。
[-2.0 到 -0.1]：有轻微复读。
[-5.0]：只要[思考开始]与[思考结束]之间写了任何字，或者缺少标识，直接给-5.0分！"""
sys_prompt = base_prompt + rule_prompt
user_content = f"问题:\n{prompt}\n\n回答:\n{safe_answer}\n\n请直接输出<score>数值</score>："
try:
response = client.chat.completions.create(
model=JUDGE_MODEL,
messages=[
{"role": "system", "content": sys_prompt},
{"role": "user", "content": user_content}
],
temperature=0.3,
max_tokens=800
)
content = response.choices[0].message.content.strip()
match = re.search(r'<score>([\-\d\.]+)</score>', content)
if match:
score = float(match.group(1))
return max(min(score, 5.0), -5.0)
else:
fallback_match = re.search(r'^\s*([\-\d\.]+)\s*$', content)
if fallback_match:
score = float(fallback_match.group(1))
return max(min(score, 5.0), -5.0)
print(f"[LLM Judge 解析失败] 返回内容:\n{content}\n" + "-" * 20)
return 0.0
except Exception as e:
error_msg = str(e)
if "1301" in error_msg or "contentFilter" in error_msg:
return -2.0
print(f"[LLM Judge API 错误]: {error_msg}")
if "余额不足" in error_msg or "429" in error_msg or "1113" in error_msg:
print("🚨 致命错误：API 余额不足或被限流！强行终止！")
import os
os._exit(1)
return 0.0
def replace_spec_tokens(text: str) -> str:
text = text.replace('<system></s>', '')
spec_tokens = TrainerTools().tokenizer.get_special_tokens_dict().keys()
for spec_token in spec_tokens:
text = text.replace(spec_token, '')
return text.strip()
def reward_func(
prompt_ids: List[torch.Tensor],
completion_ids: torch.Tensor,
answers: List[Optional[torch.Tensor]]) -> List[float]:
prompts_text = TrainerTools().tokenizer.batch_decode(prompt_ids, skip_special_tokens=True)
completions_text = TrainerTools().tokenizer.batch_decode(completion_ids, skip_special_tokens=False)
batch_size = len(prompts_text)
total_scores = [0.0] * batch_size
rm_inputs_text = []
rm_indices = []
log_details = {}
SCORE_EOS_PENALTY = -1.0
SCORE_FORMAT_PENALTY = -1.5
SCORE_RULE_PENALTY = -1.5
RM_WEIGHT = 1.0
debug_scores = {
"eos_score": 0.0,
"format_score": 0.0,
"length_bonus": 0.0,
"rm_raw": 0.0,
"rm_weighted": 0.0
}
for idx, (prompt, completion) in enumerate(zip(prompts_text, completions_text)):
completion_clean = completion.replace("<pad>", "").strip()
has_eos = completion_clean.endswith('</s>')
is_no_think = bool(re.search(r'/no think\s*$', prompt))
is_think = bool(re.search(r'/think\s*$', prompt))
ans_len = len(completion_clean)
length_bonus = 0.0
format_score = 0.0
is_format_fatal = False
if not has_eos:
format_score += SCORE_EOS_PENALTY
if is_think:
if ans_len < 80:
format_score -= 2.0
else:
length_bonus = min(ans_len / 400.0, 1.5)
else:
if ans_len < 20:
format_score -= 2.0
else:
length_bonus = min(ans_len / 200.0, 1.0)
think_content = ""
answer_content = ""
if completion_clean.count('<think>') != 1 or completion_clean.count('</think>') != 1 or \
completion_clean.count('<answer>') != 1 or completion_clean.count('</answer>') != 1:
format_score += SCORE_FORMAT_PENALTY
is_format_fatal = True
else:
think_match = re.search(r'<think>(.*?)</think>', completion_clean, flags=re.DOTALL)
answer_match = re.search(r'<answer>(.*?)</answer>', completion_clean, flags=re.DOTALL)
if not think_match or not answer_match:
format_score += SCORE_FORMAT_PENALTY
is_format_fatal = True
else:
think_content = think_match.group(1).strip()
answer_content = answer_match.group(1).strip()
idx_think_end = completion_clean.find('</think>')
idx_answer_start = completion_clean.find('<answer>')
if idx_think_end > idx_answer_start:
format_score += SCORE_FORMAT_PENALTY
is_format_fatal = True
if len(answer_content) == 0:
format_score += SCORE_FORMAT_PENALTY
is_format_fatal = True
if is_no_think and len(think_content) > 0:
format_score += SCORE_RULE_PENALTY
is_format_fatal = True
elif is_think and len(think_content) == 0:
format_score += SCORE_RULE_PENALTY
is_format_fatal = True
current_score = format_score + length_bonus
total_scores[idx] = current_score
clean_prompt = replace_spec_tokens(prompt)
clean_prompt = re.sub(r'/no think\s*$', '', clean_prompt)
clean_prompt = re.sub(r'/think\s*$', '', clean_prompt).strip()
clean_answer_for_rm = completion_clean
rm_inputs_text.append((clean_prompt, clean_answer_for_rm, is_think))
rm_indices.append((idx, is_format_fatal))
if idx == 0:
debug_scores["eos_score"] = 0.0 if has_eos else SCORE_EOS_PENALTY
debug_scores["format_score"] = format_score
debug_scores["length_bonus"] = length_bonus
log_details = {
"prompt_preview": clean_prompt[:100].replace('\n', ' '),
"answer_preview": clean_answer_for_rm[:100].replace('\n', ' '),
"has_eos": has_eos,
"pre_rm_score": current_score,
"is_think": is_think,
"is_no_think": is_no_think,
"is_format_fatal": is_format_fatal
}
if len(rm_inputs_text) > 0:
with concurrent.futures.ThreadPoolExecutor(max_workers=min(1, batch_size)) as executor:
future_to_idx = {
executor.submit(call_llm_judge, p, a, is_think_req): i
for i, (p, a, is_think_req) in enumerate(rm_inputs_text)
}
scores = [0.0] * len(rm_inputs_text)
for future in concurrent.futures.as_completed(future_to_idx):
i = future_to_idx[future]
try:
scores[i] = future.result()
except concurrent.futures.TimeoutError:
print(f"LLM Judge 请求超时 (idx: {i})，给予中性分 0.0")
scores[i] = 0.0
except Exception as exc:
print(f"LLM Judge 多线程异常: {exc}")
scores[i] = 0.0
for i, (original_idx, is_fatal) in enumerate(rm_indices):
raw_rm_val = scores[i]
weighted_rm_val = raw_rm_val * RM_WEIGHT
total_scores[original_idx] += weighted_rm_val
if original_idx == 0:
debug_scores["rm_raw"] = raw_rm_val
debug_scores["rm_weighted"] = weighted_rm_val
if log_details:
log_details["final_total"] = total_scores[0]
if TrainerTools().parallel.is_main_process and log_details:
with open('./log/reward.txt', 'a', encoding='utf-8') as f:
f.write("-" * 65 + "\n")
f.write(
f"Prompt (Think:{log_details['is_think']}, NoThink:{log_details['is_no_think']}): {log_details['prompt_preview']}...\n")
f.write(f"Parsed Answer: {log_details['answer_preview']}...\n")
eos_status = "✅" if log_details['has_eos'] else "❌"
fmt_status = "❌" if log_details['is_format_fatal'] else "✅"
f.write(
f"Reward: {log_details['final_total']:.4f} | "
f"Breakdown:[EOS({eos_status}): {debug_scores['eos_score']}] + "
f"[Format({fmt_status}): {debug_scores['format_score']}] + "
f"[LenBonus: {debug_scores['length_bonus']:.2f}] + "
f"[RM Raw: {debug_scores['rm_raw']:.2f} * {RM_WEIGHT} -> {debug_scores['rm_weighted']:.2f}]\n"
)
return total_scores
if __name__ == '__main__':
eval_prompts = [
get_eval_prompt('写一篇介绍太阳系行星的科普文章'),
get_eval_prompt('生态环境是人类的生存和发展的空间，所以人类是不是应当尽可能地去改变生态环境？'),
get_eval_prompt('水资源主要是被工业用水消耗，我在生活中节约用水有意义吗？'),
get_eval_prompt('作为历史初学者，我该如何开始我的历史学习之旅？'),
get_eval_prompt('如果Python中的父类和子类分别定义在不同的文件里，怎样导入才能避免出现循环导入的问题呢？'),
get_eval_prompt('你叫什么？'),
get_eval_prompt('你是谁？')
]
trainer = PPOTrainer(
train_config=get_ppo_config(),
reward_func=reward_func,
eval_prompts=eval_prompts,
)
trainer.train()