tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

fix mistral prompt assembly (#982)

Browse Source 
* fix mistral prompts

* fix spacing

* remove elif
This commit is contained in:
Hamel Husain
2023-12-21 08:00:55 -08:00
committed by GitHub
parent 161bcb6517
commit 7bbaac98f7
2 changed files with 108 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
src/axolotl/monkeypatch/fastchat_conversation_turns.py
View File

@@ -82,7 +82,7 @@ def get_turns( # pylint: disable=too-many-return-statements
else:
yield role + ":", ""
return
if self.sep_style == SeparatorStyle.LLAMA2:
if self.sep_style == SeparatorStyle.LLAMA2 and self.name != "mistral":
if self.system_message:
if self.messages:
# For llama, the system message is incorporated into the first human instruction
@@ -101,6 +101,28 @@ def get_turns( # pylint: disable=too-many-return-statements
else:
yield role, ""
return
if self.sep_style == SeparatorStyle.LLAMA2 and self.name == "mistral":
contains_sys_msg = False
if self.system_message:
contains_sys_msg = True
if self.messages:
# There is no clear guidance on how to handle system messages in Mistral so we just prepend it to the first human instruction seperated by a newline
first_role, first_msg = self.messages[0]
if first_role == self.roles[0]:
system_prompt = self.system_template.format(
system_message=" " + self.system_message
)
system_prompt += first_msg
self.messages.pop(0)
yield "", system_prompt
for i, (role, message) in enumerate(self.messages):
if message and i == 0 and not contains_sys_msg:
yield "", system_prompt.strip() + " " + message # if there is no system message, we need to make sure there is the a `<s> [INST]` at the beginning of the first instruction.
elif message:
yield role + " ", message
else:
yield role, ""
return
if self.sep_style == SeparatorStyle.CHATGLM:
# source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
# source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926