shen/chatgpt-on-wechat
1
0
Fork 0
mirror of https://github.com/zhayujie/chatgpt-on-wechat.git synced 2026-01-19 01:21:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

Lower Gemini's safety thresholds

Browse Source 
Gemini's default safety thresholds are set too high, resulting in frequent censorship of generated text. I have lowered the thresholds for all four safety categories according to Google's documentation.
This commit is contained in:
KurolVeko
2024-08-30 17:00:51 +08:00
committed by GitHub
parent 469524e8ae
commit ad77ad8f2b
1 changed files with 35 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
bot/gemini/google_gemini_bot.py
View File

@@ -14,6 +14,7 @@ from bridge.reply import Reply, ReplyType
from common.log import logger
from config import conf
from bot.baidu.baidu_wenxin_session import BaiduWenxinSession
from google.generativeai.types import HarmCategory, HarmBlockThreshold
# OpenAI对话模型API (可用)
@@ -38,16 +39,41 @@ class GoogleGeminiBot(Bot):
gemini_messages = self._convert_to_gemini_messages(self.filter_messages(session.messages))
genai.configure(api_key=self.api_key)
model = genai.GenerativeModel(self.model)
response = model.generate_content(gemini_messages)
reply_text = response.text
self.sessions.session_reply(reply_text, session_id)
logger.info(f"[Gemini] reply={reply_text}")
return Reply(ReplyType.TEXT, reply_text)
# 添加安全设置
safety_settings = {
HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_NONE,
HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE,
HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE,
}
# 生成回复，包含安全设置
response = model.generate_content(
gemini_messages,
safety_settings=safety_settings
)
if response.candidates and response.candidates[0].content:
reply_text = response.candidates[0].content.parts[0].text
logger.info(f"[Gemini] reply={reply_text}")
self.sessions.session_reply(reply_text, session_id)
return Reply(ReplyType.TEXT, reply_text)
else:
# 没有有效响应内容，可能内容被屏蔽，输出安全评分
logger.warning("[Gemini] No valid response generated. Checking safety ratings.")
if hasattr(response, 'candidates') and response.candidates:
for rating in response.candidates[0].safety_ratings:
logger.warning(f"Safety rating: {rating.category} - {rating.probability}")
error_message = "No valid response generated due to safety constraints."
self.sessions.session_reply(error_message, session_id)
return Reply(ReplyType.ERROR, error_message)
except Exception as e:
logger.error("[Gemini] fetch reply error, may contain unsafe content")
logger.error(e)
return Reply(ReplyType.ERROR, "invoke [Gemini] api failed!")
logger.error(f"[Gemini] Error generating response: {str(e)}", exc_info=True)
error_message = "Failed to invoke [Gemini] api!"
self.sessions.session_reply(error_message, session_id)
return Reply(ReplyType.ERROR, error_message)
def _convert_to_gemini_messages(self, messages: list):
res = []
for msg in messages: