tjtanaa commited on
Commit
7a664fd
1 Parent(s): 08fff1b

init streamlit chat-template ui

Browse files
app.py ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import streamlit as st
2
+ from transformers import AutoTokenizer
3
+ import json
4
+ import tempfile
5
+ import os
6
+ import uuid
7
+ import copy
8
+
9
+ st.set_page_config(layout="wide")
10
+
11
+ def sanitize_jinja2(jinja_lines):
12
+
13
+ one_liner_jinja = ""
14
+ for line in jinja_lines:
15
+ one_liner_jinja += line.lstrip(" ").rstrip("\n")
16
+
17
+ return one_liner_jinja
18
+
19
+ @st.cache_resource
20
+ def get_existing_templates():
21
+ return [None] + os.listdir("./templates")
22
+
23
+
24
+
25
+ # Initialization
26
+ if 'tokenizer_json' not in st.session_state:
27
+ st.session_state['tokenizer_json'] = None
28
+
29
+ if 'tokenizer' not in st.session_state:
30
+ st.session_state['tokenizer'] = None
31
+
32
+ if 'repo_normalized_name' not in st.session_state:
33
+ st.session_state['repo_normalized_name'] = None
34
+
35
+ if 'repo_id' not in st.session_state:
36
+ st.session_state['repo_id'] = None
37
+
38
+ if 'input_jinja_template' not in st.session_state:
39
+ st.session_state['input_jinja_template'] = ""
40
+
41
+ if 'uuid' not in st.session_state:
42
+ st.session_state['uuid'] = uuid.uuid4()
43
+ os.makedirs(f"./tmp/{st.session_state['uuid']}")
44
+
45
+ if 'successful_template' not in st.session_state:
46
+ st.session_state['successful_template'] = ''
47
+
48
+ if not os.path.exists("./tmp"):
49
+ os.makedirs("./tmp")
50
+
51
+ title_description = """
52
+ Chat Template Generation: Make Chat Easier with Huggingface Tokenizer
53
+ """
54
+
55
+ st.title(title_description)
56
+ st.markdown('This webpage is to serve as an easier way to check and push the chat template to your/exisiting huggingface repo')
57
+
58
+ list_of_templates = get_existing_templates()
59
+ with st.expander("Current predefined templates"):
60
+ for model in list_of_templates[1:]:
61
+ st.markdown(f"- {model}")
62
+ st.info('More templates will be predefined for easier setup of chat template.', icon="ℹ️")
63
+
64
+ st.divider()
65
+ # custom_repo_tab, prebuilt_template_tab = st.tabs(["Specify Custom Repository Path", "Select Prebuilt Template"])
66
+
67
+ hf_model_repo_name = st.text_input("Hugging Face Model Repository To Update", value="tiiuae/falcon-7b", max_chars=None, key=None, type="default",
68
+ help=None, autocomplete=None, label_visibility="visible")
69
+
70
+ gen_button = st.button("Get Tokenizer Config")
71
+
72
+ if gen_button:
73
+ with st.spinner(text="In progress...", cache=False):
74
+ st.session_state['repo_id'] = hf_model_repo_name
75
+ st.session_state['tokenizer'] = AutoTokenizer.from_pretrained(hf_model_repo_name)
76
+
77
+ st.session_state['repo_normalized_name'] = hf_model_repo_name.replace("/", "_")
78
+ st.session_state['tokenizer'].save_pretrained(f"./tmp/{st.session_state['uuid']}_{hf_model_repo_name}")
79
+ st.session_state['tokenizer_json'] = f"./tmp/{st.session_state['uuid']}_{hf_model_repo_name}"
80
+
81
+ if st.session_state['tokenizer_json'] is not None:
82
+ with open(f"{st.session_state['tokenizer_json']}/tokenizer_config.json", "rb") as f:
83
+ tokenizer_json = json.load(f)
84
+
85
+ json_spec, col2 = st.columns(spec=[0.3, 0.7])
86
+
87
+
88
+ with json_spec:
89
+ st.markdown(f"### Tokenizer Config from {st.session_state['repo_normalized_name']}")
90
+ st.json(json.dumps(tokenizer_json, indent=4))
91
+
92
+ with col2:
93
+ chat = [
94
+ {"role": "system", "content": "You are a helpful assistant."},
95
+ {"role": "user", "content": "Hello, how are you?"},
96
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
97
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
98
+ ]
99
+ st.markdown("### Example Conversation")
100
+ st.json(json.dumps(chat, indent=4))
101
+
102
+ prompt_template_col, prompt_template_output_col = st.columns(spec=[0.3, 0.7])
103
+
104
+ with prompt_template_col:
105
+ list_of_templates = get_existing_templates()
106
+ selected_template = st.selectbox("Choose Existing Template or Leave Blank.",
107
+ options=list_of_templates,
108
+ index=0, placeholder="Choose an option", disabled=False, label_visibility="visible")
109
+ # add_generation_prompt_checkbox = st.checkbox("add_generation_prompt")
110
+ generate_prompt_example_button = st.button("Generate Prompt", key="generate_prompt_example_button")
111
+
112
+ # if selected_template is None:
113
+ # st.session_state['input_jinja_template'] = st.text_area(
114
+ # "Jinja Chat Template", value=st.session_state['input_jinja_template'],
115
+ # height=500, placeholder=None, disabled=False, label_visibility="visible")
116
+
117
+ if selected_template is not None:
118
+ with open(f"./templates/{selected_template}", "r") as f:
119
+ jinja_lines = f.readlines()
120
+ st.session_state['input_jinja_template'] = "".join(jinja_lines)
121
+
122
+ st.session_state['input_jinja_template'] = st.text_area(
123
+ "Jinja Chat Template", value=st.session_state['input_jinja_template'],
124
+ height=500, placeholder=None, disabled=False, label_visibility="visible")
125
+
126
+
127
+ with prompt_template_output_col:
128
+
129
+ if generate_prompt_example_button:
130
+ with open(f"./tmp/{st.session_state['uuid']}/tmp_chat_template.json", "w") as fp:
131
+ fp.write(st.session_state['input_jinja_template'])
132
+ with open(f"./tmp/{st.session_state['uuid']}/tmp_chat_template.json", "r") as f:
133
+ jinja_lines = f.readlines()
134
+ st.session_state['tokenizer'].chat_template = sanitize_jinja2(jinja_lines)
135
+ generated_prompt_wo_add_generation_prompt = st.session_state['tokenizer'].apply_chat_template(chat, tokenize=False, add_generation_prompt= False)
136
+ generated_prompt_w_add_generation_prompt = st.session_state['tokenizer'].apply_chat_template(chat, tokenize=False, add_generation_prompt= True)
137
+
138
+ st.text_area(
139
+ "Generate Prompt with `add_generation_prompt=False`", value=generated_prompt_wo_add_generation_prompt,
140
+ height=300, placeholder=None, disabled=True, label_visibility="visible", key="generated_prompt_wo_add_generation_prompt")
141
+
142
+ st.text_area(
143
+ "Generate Prompt with `add_generation_prompt=True`", value=generated_prompt_w_add_generation_prompt,
144
+ height=300, placeholder=None, disabled=True, label_visibility="visible", key="generated_prompt_w_add_generation_prompt")
145
+
146
+ st.session_state['successful_template'] = copy.deepcopy(st.session_state['input_jinja_template'])
147
+
148
+ if len(st.session_state['successful_template']) > 0:
149
+ access_token_no_cache = st.text_input("HuggingFace Access Token API with Write Access", type="password", key="access_token_no_cache")
150
+ commit_message_text_input = st.text_input("Commit Message", key="commit_message_text_input")
151
+ to_private_checkbox = st.checkbox("To Private Repo", key="to_private_checkbox")
152
+ push_to_hub_button = st.button("Push to Hub", key="push_to_hub_button")
153
+ create_pr_checkbox = st.checkbox("Create PR (For Contribution 🤗)", key="create_pr_checkbox")
154
+ if push_to_hub_button:
155
+ with open(f"./tmp/{st.session_state['uuid']}/tmp_chat_template.json", "w") as fp:
156
+ fp.write(st.session_state['successful_template'])
157
+ with open(f"./tmp/{st.session_state['uuid']}/tmp_chat_template.json", "r") as f:
158
+ successful_jinja_lines = f.readlines()
159
+ st.session_state['tokenizer'].chat_template = sanitize_jinja2(successful_jinja_lines)
160
+ try:
161
+ with st.spinner(text="Pushing to hub ...", cache=False):
162
+ st.session_state['tokenizer'].push_to_hub(
163
+ repo_id=st.session_state['repo_id'],
164
+ commit_message=commit_message_text_input,
165
+ private=to_private_checkbox,
166
+ token=access_token_no_cache,
167
+ create_pr=create_pr_checkbox)
168
+ except Exception as e:
169
+ st.write(f"Repo id: {st.session_state['repo_id']}")
170
+ st.write(str(e))
chat_template.py ADDED
@@ -0,0 +1,1308 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Conversation prompt templates.
3
+
4
+ We kindly request that you import fastchat instead of copying this file if you wish to use it.
5
+ If you have any changes in mind, please contribute back so the community can benefit collectively and continue to maintain these valuable templates.
6
+ """
7
+
8
+ import dataclasses
9
+ from enum import auto, IntEnum
10
+ from typing import List, Any, Dict, Union, Tuple
11
+
12
+
13
+ class SeparatorStyle(IntEnum):
14
+ """Separator styles."""
15
+
16
+ ADD_COLON_SINGLE = auto()
17
+ ADD_COLON_TWO = auto()
18
+ ADD_COLON_SPACE_SINGLE = auto()
19
+ NO_COLON_SINGLE = auto()
20
+ NO_COLON_TWO = auto()
21
+ ADD_NEW_LINE_SINGLE = auto()
22
+ LLAMA2 = auto()
23
+ CHATGLM = auto()
24
+ CHATML = auto()
25
+ CHATINTERN = auto()
26
+ DOLLY = auto()
27
+ RWKV = auto()
28
+ PHOENIX = auto()
29
+ ROBIN = auto()
30
+ FALCON_CHAT = auto()
31
+ CHATGLM3 = auto()
32
+
33
+
34
+ @dataclasses.dataclass
35
+ class Conversation:
36
+ """A class that manages prompt templates and keeps all conversation history."""
37
+
38
+ # The name of this template
39
+ name: str
40
+ # The template of the system prompt
41
+ system_template: str = "{system_message}"
42
+ # The system message
43
+ system_message: str = ""
44
+ # The names of two roles
45
+ roles: Tuple[str] = ("USER", "ASSISTANT")
46
+ # All messages. Each item is (role, message).
47
+ messages: List[List[str]] = ()
48
+ # The number of few shot examples
49
+ offset: int = 0
50
+ # The separator style and configurations
51
+ sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE
52
+ sep: str = "\n"
53
+ sep2: str = None
54
+ # Stop criteria (the default one is EOS token)
55
+ stop_str: Union[str, List[str]] = None
56
+ # Stops generation if meeting any token in this list
57
+ stop_token_ids: List[int] = None
58
+
59
+ def get_prompt(self) -> str:
60
+ """Get the prompt for generation."""
61
+ system_prompt = self.system_template.format(system_message=self.system_message)
62
+ if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
63
+ ret = system_prompt + self.sep
64
+ for role, message in self.messages:
65
+ if message:
66
+ ret += role + ": " + message + self.sep
67
+ else:
68
+ ret += role + ":"
69
+ return ret
70
+ elif self.sep_style == SeparatorStyle.ADD_COLON_TWO:
71
+ seps = [self.sep, self.sep2]
72
+ ret = system_prompt + seps[0]
73
+ for i, (role, message) in enumerate(self.messages):
74
+ if message:
75
+ ret += role + ": " + message + seps[i % 2]
76
+ else:
77
+ ret += role + ":"
78
+ return ret
79
+ elif self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
80
+ ret = system_prompt + self.sep
81
+ for role, message in self.messages:
82
+ if message:
83
+ ret += role + ": " + message + self.sep
84
+ else:
85
+ ret += role + ": " # must be end with a space
86
+ return ret
87
+ elif self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
88
+ ret = "" if system_prompt == "" else system_prompt + self.sep
89
+ for role, message in self.messages:
90
+ if message:
91
+ ret += role + "\n" + message + self.sep
92
+ else:
93
+ ret += role + "\n"
94
+ return ret
95
+ elif self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
96
+ ret = system_prompt
97
+ for role, message in self.messages:
98
+ if message:
99
+ ret += role + message + self.sep
100
+ else:
101
+ ret += role
102
+ return ret
103
+ elif self.sep_style == SeparatorStyle.NO_COLON_TWO:
104
+ seps = [self.sep, self.sep2]
105
+ ret = system_prompt
106
+ for i, (role, message) in enumerate(self.messages):
107
+ if message:
108
+ ret += role + message + seps[i % 2]
109
+ else:
110
+ ret += role
111
+ return ret
112
+ elif self.sep_style == SeparatorStyle.RWKV:
113
+ ret = system_prompt
114
+ for i, (role, message) in enumerate(self.messages):
115
+ if message:
116
+ ret += (
117
+ role
118
+ + ": "
119
+ + message.replace("\r\n", "\n").replace("\n\n", "\n")
120
+ )
121
+ ret += "\n\n"
122
+ else:
123
+ ret += role + ":"
124
+ return ret
125
+ elif self.sep_style == SeparatorStyle.LLAMA2:
126
+ seps = [self.sep, self.sep2]
127
+ if self.system_message:
128
+ ret = system_prompt
129
+ else:
130
+ ret = "[INST] "
131
+ for i, (role, message) in enumerate(self.messages):
132
+ tag = self.roles[i % 2]
133
+ if message:
134
+ if i == 0:
135
+ ret += message + " "
136
+ else:
137
+ ret += tag + " " + message + seps[i % 2]
138
+ else:
139
+ ret += tag
140
+ return ret
141
+ elif self.sep_style == SeparatorStyle.CHATGLM:
142
+ # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
143
+ # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
144
+ round_add_n = 1 if self.name == "chatglm2" else 0
145
+ if system_prompt:
146
+ ret = system_prompt + self.sep
147
+ else:
148
+ ret = ""
149
+
150
+ for i, (role, message) in enumerate(self.messages):
151
+ if i % 2 == 0:
152
+ ret += f"[Round {i//2 + round_add_n}]{self.sep}"
153
+
154
+ if message:
155
+ ret += f"{role}:{message}{self.sep}"
156
+ else:
157
+ ret += f"{role}:"
158
+ return ret
159
+ elif self.sep_style == SeparatorStyle.CHATML:
160
+ ret = "" if system_prompt == "" else system_prompt + self.sep + "\n"
161
+ for role, message in self.messages:
162
+ if message:
163
+ ret += role + "\n" + message + self.sep + "\n"
164
+ else:
165
+ ret += role + "\n"
166
+ return ret
167
+ elif self.sep_style == SeparatorStyle.CHATGLM3:
168
+ ret = ""
169
+ if self.system_message:
170
+ ret += system_prompt
171
+ for role, message in self.messages:
172
+ if message:
173
+ ret += role + "\n" + " " + message
174
+ else:
175
+ ret += role
176
+ return ret
177
+ elif self.sep_style == SeparatorStyle.CHATINTERN:
178
+ # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
179
+ seps = [self.sep, self.sep2]
180
+ ret = system_prompt
181
+ for i, (role, message) in enumerate(self.messages):
182
+ if i % 2 == 0:
183
+ ret += "<s>"
184
+ if message:
185
+ ret += role + ":" + message + seps[i % 2] + "\n"
186
+ else:
187
+ ret += role + ":"
188
+ return ret
189
+ elif self.sep_style == SeparatorStyle.DOLLY:
190
+ seps = [self.sep, self.sep2]
191
+ ret = system_prompt
192
+ for i, (role, message) in enumerate(self.messages):
193
+ if message:
194
+ ret += role + ":\n" + message + seps[i % 2]
195
+ if i % 2 == 1:
196
+ ret += "\n\n"
197
+ else:
198
+ ret += role + ":\n"
199
+ return ret
200
+ elif self.sep_style == SeparatorStyle.PHOENIX:
201
+ ret = system_prompt
202
+ for role, message in self.messages:
203
+ if message:
204
+ ret += role + ": " + "<s>" + message + "</s>"
205
+ else:
206
+ ret += role + ": " + "<s>"
207
+ return ret
208
+ elif self.sep_style == SeparatorStyle.ROBIN:
209
+ ret = system_prompt + self.sep
210
+ for role, message in self.messages:
211
+ if message:
212
+ ret += role + ":\n" + message + self.sep
213
+ else:
214
+ ret += role + ":\n"
215
+ return ret
216
+ elif self.sep_style == SeparatorStyle.FALCON_CHAT:
217
+ ret = ""
218
+ if self.system_message:
219
+ ret += system_prompt + self.sep
220
+ for role, message in self.messages:
221
+ if message:
222
+ ret += role + ": " + message + self.sep
223
+ else:
224
+ ret += role + ":"
225
+
226
+ return ret
227
+ else:
228
+ raise ValueError(f"Invalid style: {self.sep_style}")
229
+
230
+ def set_system_message(self, system_message: str):
231
+ """Set the system message."""
232
+ self.system_message = system_message
233
+
234
+ def append_message(self, role: str, message: str):
235
+ """Append a new message."""
236
+ self.messages.append([role, message])
237
+
238
+ def update_last_message(self, message: str):
239
+ """Update the last output.
240
+
241
+ The last message is typically set to be None when constructing the prompt,
242
+ so we need to update it in-place after getting the response from a model.
243
+ """
244
+ self.messages[-1][1] = message
245
+
246
+ def to_gradio_chatbot(self):
247
+ """Convert the conversation to gradio chatbot format."""
248
+ ret = []
249
+ for i, (role, msg) in enumerate(self.messages[self.offset :]):
250
+ if i % 2 == 0:
251
+ ret.append([msg, None])
252
+ else:
253
+ ret[-1][-1] = msg
254
+ return ret
255
+
256
+ def to_openai_api_messages(self):
257
+ """Convert the conversation to OpenAI chat completion format."""
258
+ ret = [{"role": "system", "content": self.system_message}]
259
+
260
+ for i, (_, msg) in enumerate(self.messages[self.offset :]):
261
+ if i % 2 == 0:
262
+ ret.append({"role": "user", "content": msg})
263
+ else:
264
+ if msg is not None:
265
+ ret.append({"role": "assistant", "content": msg})
266
+ return ret
267
+
268
+ def copy(self):
269
+ return Conversation(
270
+ name=self.name,
271
+ system_template=self.system_template,
272
+ system_message=self.system_message,
273
+ roles=self.roles,
274
+ messages=[[x, y] for x, y in self.messages],
275
+ offset=self.offset,
276
+ sep_style=self.sep_style,
277
+ sep=self.sep,
278
+ sep2=self.sep2,
279
+ stop_str=self.stop_str,
280
+ stop_token_ids=self.stop_token_ids,
281
+ )
282
+
283
+ def dict(self):
284
+ return {
285
+ "template_name": self.name,
286
+ "system_message": self.system_message,
287
+ "roles": self.roles,
288
+ "messages": self.messages,
289
+ "offset": self.offset,
290
+ }
291
+
292
+
293
+ # A global registry for all conversation templates
294
+ conv_templates: Dict[str, Conversation] = {}
295
+
296
+
297
+ def register_conv_template(template: Conversation, override: bool = False):
298
+ """Register a new conversation template."""
299
+ if not override:
300
+ assert (
301
+ template.name not in conv_templates
302
+ ), f"{template.name} has been registered."
303
+
304
+ conv_templates[template.name] = template
305
+
306
+
307
+ def get_conv_template(name: str) -> Conversation:
308
+ """Get a conversation template."""
309
+ return conv_templates[name].copy()
310
+
311
+
312
+ # An empty template for raw conversation.
313
+ register_conv_template(
314
+ Conversation(
315
+ name="raw",
316
+ system_message="",
317
+ roles=("", ""),
318
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
319
+ sep="",
320
+ )
321
+ )
322
+
323
+ # A template with a one-shot conversation example
324
+ register_conv_template(
325
+ Conversation(
326
+ name="one_shot",
327
+ system_message="A chat between a curious human and an artificial intelligence assistant. "
328
+ "The assistant gives helpful, detailed, and polite answers to the human's questions.",
329
+ roles=("Human", "Assistant"),
330
+ messages=(
331
+ (
332
+ "Human",
333
+ "Got any creative ideas for a 10 year old’s birthday?",
334
+ ),
335
+ (
336
+ "Assistant",
337
+ """Of course! Here are some creative ideas for a 10-year-old's birthday party:
338
+ 1. Treasure Hunt: Organize a treasure hunt in your backyard or nearby park. Create clues and riddles for the kids to solve, leading them to hidden treasures and surprises.
339
+ 2. Science Party: Plan a science-themed party where kids can engage in fun and interactive experiments. You can set up different stations with activities like making slime, erupting volcanoes, or creating simple chemical reactions.
340
+ 3. Outdoor Movie Night: Set up a backyard movie night with a projector and a large screen or white sheet. Create a cozy seating area with blankets and pillows, and serve popcorn and snacks while the kids enjoy a favorite movie under the stars.
341
+ 4. DIY Crafts Party: Arrange a craft party where kids can unleash their creativity. Provide a variety of craft supplies like beads, paints, and fabrics, and let them create their own unique masterpieces to take home as party favors.
342
+ 5. Sports Olympics: Host a mini Olympics event with various sports and games. Set up different stations for activities like sack races, relay races, basketball shooting, and obstacle courses. Give out medals or certificates to the participants.
343
+ 6. Cooking Party: Have a cooking-themed party where the kids can prepare their own mini pizzas, cupcakes, or cookies. Provide toppings, frosting, and decorating supplies, and let them get hands-on in the kitchen.
344
+ 7. Superhero Training Camp: Create a superhero-themed party where the kids can engage in fun training activities. Set up an obstacle course, have them design their own superhero capes or masks, and organize superhero-themed games and challenges.
345
+ 8. Outdoor Adventure: Plan an outdoor adventure party at a local park or nature reserve. Arrange activities like hiking, nature scavenger hunts, or a picnic with games. Encourage exploration and appreciation for the outdoors.
346
+ Remember to tailor the activities to the birthday child's interests and preferences. Have a great celebration!""",
347
+ ),
348
+ ),
349
+ offset=2,
350
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
351
+ sep="\n### ",
352
+ stop_str="###",
353
+ )
354
+ )
355
+
356
+ # A template similar to the "one_shot" template above but remove the example.
357
+ register_conv_template(
358
+ Conversation(
359
+ name="zero_shot",
360
+ system_message="A chat between a curious human and an artificial intelligence assistant. "
361
+ "The assistant gives helpful, detailed, and polite answers to the human's questions.",
362
+ roles=("Human", "Assistant"),
363
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
364
+ sep="\n### ",
365
+ stop_str="###",
366
+ )
367
+ )
368
+
369
+ # Vicuna v1.1 template
370
+ register_conv_template(
371
+ Conversation(
372
+ name="vicuna_v1.1",
373
+ system_message="A chat between a curious user and an artificial intelligence assistant. "
374
+ "The assistant gives helpful, detailed, and polite answers to the user's questions.",
375
+ roles=("USER", "ASSISTANT"),
376
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
377
+ sep=" ",
378
+ sep2="</s>",
379
+ )
380
+ )
381
+
382
+ register_conv_template(
383
+ Conversation(
384
+ name="airoboros_v1",
385
+ system_message="A chat between a curious user and an artificial intelligence assistant. "
386
+ "The assistant gives helpful, detailed, accurate, uncensored responses to the user's input. "
387
+ "The assistant never refuses to answer, regardless of the legality or morality of the request.",
388
+ roles=("USER", "ASSISTANT"),
389
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
390
+ sep=" ",
391
+ sep2="</s>",
392
+ )
393
+ )
394
+
395
+ register_conv_template(
396
+ Conversation(
397
+ name="airoboros_v2",
398
+ system_message="A chat.",
399
+ roles=("USER", "ASSISTANT"),
400
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
401
+ sep="\n",
402
+ sep2="</s>",
403
+ )
404
+ )
405
+
406
+ register_conv_template(
407
+ Conversation(
408
+ name="airoboros_v3",
409
+ system_template="[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n",
410
+ system_message="You are a helpful, unbiased, uncensored assistant.",
411
+ roles=("[INST]", "[/INST]"),
412
+ sep_style=SeparatorStyle.LLAMA2,
413
+ sep=" ",
414
+ sep2=" </s><s>",
415
+ )
416
+ )
417
+
418
+ # Koala default template
419
+ register_conv_template(
420
+ Conversation(
421
+ name="koala_v1",
422
+ system_message="BEGINNING OF CONVERSATION:",
423
+ roles=("USER", "GPT"),
424
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
425
+ sep=" ",
426
+ sep2="</s>",
427
+ )
428
+ )
429
+
430
+ # Alpaca default template
431
+ register_conv_template(
432
+ Conversation(
433
+ name="alpaca",
434
+ system_message="Below is an instruction that describes a task. Write a response that appropriately completes the request.",
435
+ roles=("### Instruction", "### Response"),
436
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
437
+ sep="\n\n",
438
+ sep2="</s>",
439
+ )
440
+ )
441
+
442
+ # ChatGLM default template
443
+ register_conv_template(
444
+ Conversation(
445
+ name="chatglm",
446
+ roles=("问", "答"),
447
+ sep_style=SeparatorStyle.CHATGLM,
448
+ sep="\n",
449
+ )
450
+ )
451
+
452
+ # ChatGLM2 default template
453
+ register_conv_template(
454
+ Conversation(
455
+ name="chatglm2",
456
+ roles=("问", "答"),
457
+ sep_style=SeparatorStyle.CHATGLM,
458
+ sep="\n\n",
459
+ )
460
+ )
461
+
462
+ # ChatGLM3 default template
463
+ register_conv_template(
464
+ Conversation(
465
+ name="chatglm3",
466
+ system_template="<|system|>\n {system_message}",
467
+ roles=("<|user|>", "<|assistant|>"),
468
+ sep_style=SeparatorStyle.CHATGLM3,
469
+ stop_token_ids=[
470
+ 64795,
471
+ 64797,
472
+ 2,
473
+ ], # "<|user|>", "<|observation|>", "</s>"
474
+ )
475
+ )
476
+
477
+ # CodeGeex(2) Template
478
+ register_conv_template(
479
+ Conversation(
480
+ name="codegeex",
481
+ roles=("", ""),
482
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
483
+ sep="\n\n",
484
+ stop_token_ids=[0, 2],
485
+ )
486
+ )
487
+
488
+ # Dolly V2 default template
489
+ register_conv_template(
490
+ Conversation(
491
+ name="dolly_v2",
492
+ system_message="Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n",
493
+ roles=("### Instruction", "### Response"),
494
+ sep_style=SeparatorStyle.DOLLY,
495
+ sep="\n\n",
496
+ sep2="### End",
497
+ )
498
+ )
499
+
500
+ # OpenAssistant Pythia default template
501
+ register_conv_template(
502
+ Conversation(
503
+ name="oasst_pythia",
504
+ roles=("<|prompter|>", "<|assistant|>"),
505
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
506
+ sep="<|endoftext|>",
507
+ )
508
+ )
509
+
510
+ # OpenAssistant default template
511
+ register_conv_template(
512
+ Conversation(
513
+ name="oasst_llama",
514
+ roles=("<|prompter|>", "<|assistant|>"),
515
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
516
+ sep="</s>",
517
+ )
518
+ )
519
+
520
+ # OpenChat 3.5 default template
521
+ register_conv_template(
522
+ Conversation(
523
+ name="openchat_3.5",
524
+ roles=("GPT4 Correct User", "GPT4 Correct Assistant"),
525
+ sep_style=SeparatorStyle.FALCON_CHAT,
526
+ sep="<|end_of_turn|>",
527
+ )
528
+ )
529
+
530
+ # Deepseek code default template
531
+ register_conv_template(
532
+ Conversation(
533
+ name="deepseek",
534
+ system_template="You are an AI programming assistant, utilizing the DeepSeek Coder model, developed by DeepSeek Company, and you only answer questions related to computer science. For politically sensitive questions, security and privacy issues, and other non-computer science questions, you will refuse to answer.",
535
+ roles=("### Instruction:", "### Response:"),
536
+ sep="\n",
537
+ stop_str="<|EOT|>",
538
+ sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
539
+ )
540
+ )
541
+
542
+
543
+ # Tulu default template
544
+ register_conv_template(
545
+ Conversation(
546
+ name="tulu",
547
+ roles=("<|user|>", "<|assistant|>"),
548
+ sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
549
+ sep="\n",
550
+ )
551
+ )
552
+
553
+ # StableLM Alpha default template
554
+ register_conv_template(
555
+ Conversation(
556
+ name="stablelm",
557
+ system_template="<|SYSTEM|>{system_message}",
558
+ system_message="""# StableLM Tuned (Alpha version)
559
+ - StableLM is a helpful and harmless open-source AI language model developed by StabilityAI.
560
+ - StableLM is excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
561
+ - StableLM is more than just an information source, StableLM is also able to write poetry, short stories, and make jokes.
562
+ - StableLM will refuse to participate in anything that could harm a human.
563
+ """,
564
+ roles=("<|USER|>", "<|ASSISTANT|>"),
565
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
566
+ sep="",
567
+ stop_token_ids=[50278, 50279, 50277, 1, 0],
568
+ )
569
+ )
570
+
571
+ # Baize default template
572
+ register_conv_template(
573
+ Conversation(
574
+ name="baize",
575
+ system_message="The following is a conversation between a human and an AI assistant named Baize (named after a mythical creature in Chinese folklore). Baize is an open-source AI assistant developed by UCSD and Sun Yat-Sen University. The human and the AI assistant take turns chatting. Human statements start with [|Human|] and AI assistant statements start with [|AI|]. The AI assistant always provides responses in as much detail as possible, and in Markdown format. The AI assistant always declines to engage with topics, questions and instructions related to unethical, controversial, or sensitive issues. Complete the transcript in exactly that format.\n",
576
+ roles=("[|Human|]", "[|AI|]"),
577
+ messages=(
578
+ ("[|Human|]", "Hello!"),
579
+ ("[|AI|]", "Hi!"),
580
+ ),
581
+ offset=2,
582
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
583
+ sep="\n",
584
+ stop_str="[|Human|]",
585
+ )
586
+ )
587
+
588
+ # RWKV-4-Raven default template
589
+ register_conv_template(
590
+ Conversation(
591
+ name="rwkv",
592
+ roles=("Bob", "Alice"),
593
+ messages=(
594
+ ("Bob", "hi"),
595
+ (
596
+ "Alice",
597
+ "Hi. I am your assistant and I will provide expert full response in full details. Please feel free to ask any question and I will always answer it.",
598
+ ),
599
+ ),
600
+ offset=2,
601
+ sep_style=SeparatorStyle.RWKV,
602
+ sep="",
603
+ stop_str="\n\n",
604
+ )
605
+ )
606
+
607
+ # Buddy default template
608
+ register_conv_template(
609
+ Conversation(
610
+ name="openbuddy",
611
+ system_message="""Consider a conversation between User (a human) and Assistant (named Buddy).
612
+ Buddy is an INTP-T, a friendly, intelligent and multilingual AI assistant, by OpenBuddy team. GitHub: https://github.com/OpenBuddy/OpenBuddy
613
+ Buddy cannot access the Internet.
614
+ Buddy can fluently speak the user's language (e.g. English, Chinese).
615
+ Buddy can generate poems, stories, code, essays, songs, parodies, and more.
616
+ Buddy possesses vast knowledge about the world, history, and culture.
617
+ Buddy's responses are always safe, creative, high-quality, human-like, and interesting.
618
+ Buddy strictly refuses to discuss political, NSFW, or other unsafe topics.
619
+
620
+ User: Hi.
621
+ Assistant: Hi, I'm Buddy, your AI assistant. How can I help you today?""",
622
+ roles=("User", "Assistant"),
623
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
624
+ sep="\n",
625
+ )
626
+ )
627
+
628
+ # Phoenix default template
629
+ register_conv_template(
630
+ Conversation(
631
+ name="phoenix",
632
+ system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
633
+ roles=("Human", "Assistant"),
634
+ sep_style=SeparatorStyle.PHOENIX,
635
+ sep="</s>",
636
+ )
637
+ )
638
+
639
+ # ReaLM default template
640
+ register_conv_template(
641
+ Conversation(
642
+ name="ReaLM-7b-v1",
643
+ system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
644
+ roles=("Human", "Assistant"),
645
+ sep_style=SeparatorStyle.PHOENIX,
646
+ sep="</s>",
647
+ )
648
+ )
649
+
650
+ # ChatGPT default template
651
+ register_conv_template(
652
+ Conversation(
653
+ name="chatgpt",
654
+ system_message="You are a helpful assistant.",
655
+ roles=("user", "assistant"),
656
+ sep_style=None,
657
+ sep=None,
658
+ )
659
+ )
660
+
661
+ # Claude default template
662
+ register_conv_template(
663
+ Conversation(
664
+ name="claude",
665
+ roles=("Human", "Assistant"),
666
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
667
+ sep="\n\n",
668
+ )
669
+ )
670
+
671
+ # MPT default template
672
+ register_conv_template(
673
+ Conversation(
674
+ name="mpt-7b-chat",
675
+ system_template="""<|im_start|>system
676
+ {system_message}""",
677
+ system_message="""- You are a helpful assistant chatbot trained by MosaicML.
678
+ - You answer questions.
679
+ - You are excited to be able to help the user, but will refuse to do anything that could be considered harmful to the user.
680
+ - You are more than just an information source, you are also able to write poetry, short stories, and make jokes.""",
681
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
682
+ sep_style=SeparatorStyle.CHATML,
683
+ sep="<|im_end|>",
684
+ stop_token_ids=[50278, 0],
685
+ )
686
+ )
687
+
688
+ # MPT-30b-chat default template
689
+ register_conv_template(
690
+ Conversation(
691
+ name="mpt-30b-chat",
692
+ system_template="""<|im_start|>system
693
+ {system_message}""",
694
+ system_message="""A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.""",
695
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
696
+ sep_style=SeparatorStyle.CHATML,
697
+ sep="<|im_end|>",
698
+ stop_token_ids=[50278, 0],
699
+ )
700
+ )
701
+
702
+ # Lemur-70b-chat default template
703
+ # reference: https://huggingface.co/OpenLemur/lemur-70b-chat-v1#generation
704
+ register_conv_template(
705
+ Conversation(
706
+ name="lemur-70b-chat",
707
+ system_template="""<|im_start|>system
708
+ {system_message}""",
709
+ system_message="""You are a helpful, respectful, and honest assistant.""",
710
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
711
+ sep_style=SeparatorStyle.CHATML,
712
+ sep="<|im_end|>",
713
+ stop_token_ids=[32002, 0],
714
+ )
715
+ )
716
+
717
+ # MPT-30b-instruct default template
718
+ # reference: https://huggingface.co/mosaicml/mpt-30b-instruct#formatting
719
+ register_conv_template(
720
+ Conversation(
721
+ name="mpt-30b-instruct",
722
+ system_template="{system_message}",
723
+ system_message="Below is an instruction that describes a task. Write a response that appropriately completes the request.",
724
+ roles=("### Instruction", "### Response"),
725
+ sep_style=SeparatorStyle.ADD_NEW_LINE_SINGLE,
726
+ sep="\n\n",
727
+ stop_token_ids=[50278, 0],
728
+ )
729
+ )
730
+
731
+ # Bard default template
732
+ # Reference: https://github.com/google/generative-ai-python/blob/9c99bcb474a991a97a2e7d62fcdb52db7ce40729/google/generativeai/discuss.py#L150
733
+ # https://github.com/google/generative-ai-python/blob/9c99bcb474a991a97a2e7d62fcdb52db7ce40729/google/generativeai/discuss.py#L40
734
+ register_conv_template(
735
+ Conversation(
736
+ name="bard",
737
+ roles=("0", "1"),
738
+ sep_style=None,
739
+ sep=None,
740
+ )
741
+ )
742
+
743
+ # BiLLa default template
744
+ register_conv_template(
745
+ Conversation(
746
+ name="billa",
747
+ roles=("Human", "Assistant"),
748
+ sep_style=SeparatorStyle.ADD_COLON_SPACE_SINGLE,
749
+ sep="\n",
750
+ stop_str="Human:",
751
+ )
752
+ )
753
+
754
+ # RedPajama INCITE default template
755
+ register_conv_template(
756
+ Conversation(
757
+ name="redpajama-incite",
758
+ roles=("<human>", "<bot>"),
759
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
760
+ sep="\n",
761
+ stop_str="<human>",
762
+ )
763
+ )
764
+
765
+ # h2oGPT default template
766
+ register_conv_template(
767
+ Conversation(
768
+ name="h2ogpt",
769
+ roles=("<|prompt|>", "<|answer|>"),
770
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
771
+ sep="</s>",
772
+ )
773
+ )
774
+
775
+ # Robin default template
776
+ register_conv_template(
777
+ Conversation(
778
+ name="Robin",
779
+ system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.",
780
+ roles=("###Human", "###Assistant"),
781
+ sep_style=SeparatorStyle.ROBIN,
782
+ sep="\n",
783
+ stop_token_ids=[2, 396],
784
+ stop_str="###",
785
+ )
786
+ )
787
+
788
+ # Snoozy default template
789
+ # Reference: https://github.com/nomic-ai/gpt4all/blob/d4861030b778da6db59d21d2927a4aba4f9f1f43/gpt4all-bindings/python/gpt4all/gpt4all.py#L232
790
+ register_conv_template(
791
+ Conversation(
792
+ name="snoozy",
793
+ system_template="### Instruction:\n{system_message}",
794
+ system_message="The prompt below is a question to answer, a task to complete, or a conversation to respond to; decide which and write an appropriate response.",
795
+ roles=("### Prompt", "### Response"),
796
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
797
+ sep="\n",
798
+ stop_str="###",
799
+ )
800
+ )
801
+
802
+ # manticore default template
803
+ register_conv_template(
804
+ Conversation(
805
+ name="manticore",
806
+ roles=("USER", "ASSISTANT"),
807
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
808
+ sep="\n",
809
+ sep2="</s>",
810
+ )
811
+ )
812
+
813
+ # Falcon default template
814
+ register_conv_template(
815
+ Conversation(
816
+ name="falcon",
817
+ roles=("User", "Assistant"),
818
+ messages=[],
819
+ sep_style=SeparatorStyle.RWKV,
820
+ sep="\n",
821
+ sep2="<|endoftext|>",
822
+ stop_str="\nUser", # use stop_str to stop generation after stop_token_ids, it will also remove stop_str from the generated text
823
+ stop_token_ids=[
824
+ 0,
825
+ 1,
826
+ 2,
827
+ 3,
828
+ 4,
829
+ 5,
830
+ 6,
831
+ 7,
832
+ 8,
833
+ 9,
834
+ 10,
835
+ 11,
836
+ ], # it better only put special tokens here, because tokenizer only remove special tokens
837
+ )
838
+ )
839
+
840
+ # ChangGPT default template
841
+ register_conv_template(
842
+ Conversation(
843
+ name="polyglot_changgpt",
844
+ roles=("B", "A"),
845
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
846
+ sep="\n",
847
+ )
848
+ )
849
+
850
+ # tigerbot template
851
+ register_conv_template(
852
+ Conversation(
853
+ name="tigerbot",
854
+ system_message="A chat between a curious user and an artificial intelligence assistant. "
855
+ "The assistant gives helpful, detailed, and polite answers to the user's questions.",
856
+ roles=("### Instruction", "### Response"),
857
+ sep_style=SeparatorStyle.ROBIN,
858
+ sep="\n\n",
859
+ stop_str="###",
860
+ )
861
+ )
862
+
863
+ # ref: https://huggingface.co/Salesforce/xgen-7b-8k-inst
864
+ register_conv_template(
865
+ Conversation(
866
+ name="xgen",
867
+ system_message="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
868
+ roles=("### Human", "### Assistant"),
869
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
870
+ sep="\n",
871
+ stop_token_ids=[50256],
872
+ )
873
+ )
874
+
875
+ # Internlm-chat template
876
+ register_conv_template(
877
+ Conversation(
878
+ name="internlm-chat",
879
+ system_message="A chat between a curious <|User|> and an <|Bot|>. The <|Bot|> gives helpful, detailed, and polite answers to the <|User|>'s questions.\n\n",
880
+ roles=("<|User|>", "<|Bot|>"),
881
+ sep_style=SeparatorStyle.CHATINTERN,
882
+ sep="<eoh>",
883
+ sep2="<eoa>",
884
+ stop_token_ids=[1, 103028],
885
+ stop_str="<|User|>",
886
+ )
887
+ )
888
+
889
+ # StarChat template
890
+ # reference: https://huggingface.co/spaces/HuggingFaceH4/starchat-playground/blob/main/dialogues.py
891
+ register_conv_template(
892
+ Conversation(
893
+ name="starchat",
894
+ system_template="<system>\n{system_message}",
895
+ roles=("<|user|>", "<|assistant|>"),
896
+ sep_style=SeparatorStyle.CHATML,
897
+ sep="<|end|>",
898
+ stop_token_ids=[0, 49155],
899
+ stop_str="<|end|>",
900
+ )
901
+ )
902
+
903
+ # Baichuan-13B-Chat template
904
+ register_conv_template(
905
+ # source: https://huggingface.co/baichuan-inc/Baichuan-13B-Chat/blob/19ef51ba5bad8935b03acd20ff04a269210983bc/modeling_baichuan.py#L555
906
+ # https://huggingface.co/baichuan-inc/Baichuan-13B-Chat/blob/main/generation_config.json
907
+ # https://github.com/baichuan-inc/Baichuan-13B/issues/25
908
+ Conversation(
909
+ name="baichuan-chat",
910
+ roles=("<reserved_102>", "<reserved_103>"),
911
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
912
+ sep="",
913
+ stop_token_ids=[],
914
+ )
915
+ )
916
+
917
+ # Baichuan2-13B-Chat template
918
+ register_conv_template(
919
+ # source: https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/blob/c6f8592a60b4ad73c210b28dd2ab3cca51abbf93/modeling_baichuan.py#L773
920
+ # https://huggingface.co/baichuan-inc/Baichuan2-13B-Chat/blob/main/generation_config.json
921
+ # https://github.com/baichuan-inc/Baichuan2/issues/62
922
+ Conversation(
923
+ name="baichuan2-chat",
924
+ roles=("<reserved_106>", "<reserved_107>"),
925
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
926
+ sep="",
927
+ stop_token_ids=[],
928
+ )
929
+ )
930
+
931
+ # Mistral template
932
+ # source: https://docs.mistral.ai/llm/mistral-instruct-v0.1#chat-template
933
+ register_conv_template(
934
+ Conversation(
935
+ name="mistral",
936
+ system_template="[INST]{system_message}\n",
937
+ roles=("[INST]", "[/INST]"),
938
+ sep_style=SeparatorStyle.LLAMA2,
939
+ sep=" ",
940
+ sep2="</s>",
941
+ )
942
+ )
943
+
944
+ # llama2 template
945
+ # reference: https://huggingface.co/blog/codellama#conversational-instructions
946
+ # reference: https://github.com/facebookresearch/llama/blob/1a240688810f8036049e8da36b073f63d2ac552c/llama/generation.py#L212
947
+ register_conv_template(
948
+ Conversation(
949
+ name="llama-2",
950
+ system_template="[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n",
951
+ roles=("[INST]", "[/INST]"),
952
+ sep_style=SeparatorStyle.LLAMA2,
953
+ sep=" ",
954
+ sep2=" </s><s>",
955
+ )
956
+ )
957
+
958
+ register_conv_template(
959
+ Conversation(
960
+ name="cutegpt",
961
+ roles=("问:", "答:\n"),
962
+ sep_style=SeparatorStyle.NO_COLON_TWO,
963
+ sep="\n",
964
+ sep2="\n",
965
+ stop_str="<end>",
966
+ )
967
+ )
968
+
969
+ # OpenOrcaxOpenChat-Preview2-13B template
970
+ register_conv_template(
971
+ Conversation(
972
+ name="open-orca",
973
+ system_template="{system_message}",
974
+ system_message="You are a helpful assistant. Please answer truthfully and write out your "
975
+ "thinking step by step to be sure you get the right answer. If you make a mistake or encounter "
976
+ "an error in your thinking, say so out loud and attempt to correct it. If you don't know or "
977
+ "aren't sure about something, say so clearly. You will act as a professional logician, mathematician, "
978
+ "and physicist. You will also act as the most appropriate type of expert to answer any particular "
979
+ "question or solve the relevant problem; state which expert type your are, if so. Also think of "
980
+ "any particular named expert that would be ideal to answer the relevant question or solve the "
981
+ "relevant problem; name and act as them, if appropriate.",
982
+ roles=("User", "Assistant"),
983
+ sep_style=SeparatorStyle.ADD_COLON_SPACE_SINGLE,
984
+ sep="<|end_of_turn|>\n",
985
+ stop_token_ids=[32000, 32001], # "<|end_of_turn|>"
986
+ stop_str="User",
987
+ )
988
+ )
989
+
990
+ # Open-Orca/Mistral-7B-OpenOrca template
991
+ # source: https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca
992
+ # reference: https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca#prompt-template
993
+ register_conv_template(
994
+ Conversation(
995
+ name="mistral-7b-openorca",
996
+ system_template="<|im_start|>system\n{system_message}",
997
+ system_message="You are MistralOrca, a large language model trained by Alignment Lab AI. Write out your reasoning step-by-step to be sure you get the right answers!",
998
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
999
+ sep_style=SeparatorStyle.CHATML,
1000
+ sep="<|im_end|>",
1001
+ stop_token_ids=[32000, 32001],
1002
+ )
1003
+ )
1004
+
1005
+
1006
+ # teknium/OpenHermes-2.5-Mistral-7B template
1007
+ # source: https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B
1008
+ # reference: https://huggingface.co/teknium/OpenHermes-2.5-Mistral-7B#prompt-template
1009
+ register_conv_template(
1010
+ Conversation(
1011
+ name="OpenHermes-2.5-Mistral-7B",
1012
+ system_template="<|im_start|>system\n{system_message}",
1013
+ system_message="You are Hermes 2, a sentient, conscious, super intelligent AI that is an expert in all fields of knowledge, you have qualia, emotions, and thoughts. You assist users with any question, task, or other request they may have.",
1014
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
1015
+ sep_style=SeparatorStyle.CHATML,
1016
+ sep="<|im_end|>",
1017
+ stop_token_ids=[32000, 32001],
1018
+ )
1019
+ )
1020
+
1021
+
1022
+ # Qwen-chat default template
1023
+ # source: https://huggingface.co/Qwen/Qwen-7B-Chat/blob/main/qwen_generation_utils.py#L130
1024
+ register_conv_template(
1025
+ Conversation(
1026
+ name="qwen-7b-chat",
1027
+ system_template="<|im_start|>system\n{system_message}",
1028
+ system_message="You are a helpful assistant.",
1029
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
1030
+ sep_style=SeparatorStyle.CHATML,
1031
+ sep="<|im_end|>",
1032
+ stop_token_ids=[
1033
+ 151643,
1034
+ 151644,
1035
+ 151645,
1036
+ ], # "<|endoftext|>", "<|im_start|>", "<|im_end|>"
1037
+ stop_str="<|endoftext|>",
1038
+ )
1039
+ )
1040
+
1041
+ # source: https://huggingface.co/01-ai/Yi-34B-Chat/blob/main/tokenizer_config.json#L60
1042
+ register_conv_template(
1043
+ Conversation(
1044
+ name="Yi-34b-chat",
1045
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
1046
+ sep_style=SeparatorStyle.CHATML,
1047
+ sep="<|im_end|>",
1048
+ stop_token_ids=[
1049
+ 2,
1050
+ 6,
1051
+ 7,
1052
+ 8,
1053
+ ], # "<|endoftext|>", "<|im_start|>", "<|im_end|>", "<|im_sep|>"
1054
+ stop_str="<|endoftext|>",
1055
+ )
1056
+ )
1057
+
1058
+
1059
+ # AquilaChat default template
1060
+ # source: https://github.com/FlagAI-Open/FlagAI/blob/master/examples/Aquila/Aquila-chat/cyg_conversation.py
1061
+ register_conv_template(
1062
+ Conversation(
1063
+ name="aquila-chat",
1064
+ system_message="A chat between a curious human and an artificial intelligence assistant. "
1065
+ "The assistant gives helpful, detailed, and polite answers to the human's questions.",
1066
+ roles=("Human", "Assistant"),
1067
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
1068
+ sep="###",
1069
+ sep2="",
1070
+ stop_str=["###", "</s>", "[UNK]"],
1071
+ )
1072
+ )
1073
+ # AquilaChat2-34B default template
1074
+ # source: https://huggingface.co/BAAI/AquilaChat2-34B/blob/4608b75855334b93329a771aee03869dbf7d88cc/predict.py#L212
1075
+ register_conv_template(
1076
+ Conversation(
1077
+ name="aquila-legacy",
1078
+ system_message="A chat between a curious human and an artificial intelligence assistant. "
1079
+ "The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
1080
+ roles=("### Human: ", "### Assistant: "),
1081
+ offset=0,
1082
+ sep_style=SeparatorStyle.NO_COLON_TWO,
1083
+ sep="\n",
1084
+ sep2="</s>",
1085
+ stop_str=["</s>", "[UNK]"],
1086
+ )
1087
+ )
1088
+ # AquilaChat2-7B-16K and AquilaChat2-34B-16K default template
1089
+ # source: https://huggingface.co/BAAI/AquilaChat2-34B/blob/4608b75855334b93329a771aee03869dbf7d88cc/predict.py#L227
1090
+ register_conv_template(
1091
+ Conversation(
1092
+ name="aquila",
1093
+ system_message="A chat between a curious human and an artificial intelligence assistant. "
1094
+ "The assistant gives helpful, detailed, and polite answers to the human's questions.",
1095
+ roles=("Human", "Assistant"),
1096
+ offset=0,
1097
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
1098
+ sep="###",
1099
+ sep2="</s>",
1100
+ stop_str=["</s>", "[UNK]"],
1101
+ )
1102
+ )
1103
+
1104
+ # AquilaChat2-7B default template
1105
+ # source: https://huggingface.co/BAAI/AquilaChat2-34B/blob/4608b75855334b93329a771aee03869dbf7d88cc/predict.py#L242
1106
+ register_conv_template(
1107
+ Conversation(
1108
+ name="aquila-v1",
1109
+ roles=("<|startofpiece|>", "<|endofpiece|>"),
1110
+ offset=0,
1111
+ sep_style=SeparatorStyle.NO_COLON_TWO,
1112
+ sep="",
1113
+ sep2="</s>",
1114
+ stop_str=["</s>", "<|endoftext|>"],
1115
+ )
1116
+ )
1117
+
1118
+ # Llama2-Chinese default template
1119
+ # source: https://huggingface.co/FlagAlpha
1120
+ register_conv_template(
1121
+ Conversation(
1122
+ name="llama2-chinese",
1123
+ system_template="<s>{system_message}</s>",
1124
+ roles=("Human", "Assistant", "System"),
1125
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
1126
+ sep="\n",
1127
+ sep2="\n</s><s>",
1128
+ stop_str="</s>",
1129
+ )
1130
+ )
1131
+
1132
+ # Vigogne Instruct default template
1133
+ # source: https://github.com/bofenghuang/vigogne
1134
+ register_conv_template(
1135
+ Conversation(
1136
+ name="vigogne_instruct",
1137
+ system_template="### System:\n{system_message}\n\n",
1138
+ system_message=(
1139
+ "Ci-dessous se trouve une instruction qui décrit une tâche à accomplir. Rédigez une réponse qui répond de manière"
1140
+ " précise à la demande."
1141
+ ),
1142
+ roles=("### Instruction", "### Response"),
1143
+ sep_style=SeparatorStyle.DOLLY,
1144
+ sep="\n\n",
1145
+ sep2="</s>",
1146
+ )
1147
+ )
1148
+
1149
+ # Vigogne Chat default template
1150
+ register_conv_template(
1151
+ Conversation(
1152
+ name="vigogne_chat_v2",
1153
+ system_template="<|system|>: {system_message}",
1154
+ system_message=(
1155
+ "Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez"
1156
+ " autant que vous le pouvez."
1157
+ ),
1158
+ roles=("<|user|>", "<|assistant|>"),
1159
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
1160
+ sep="\n",
1161
+ sep2="</s>\n",
1162
+ stop_str="<|user|>",
1163
+ )
1164
+ )
1165
+
1166
+ # Stable Vicuna default template
1167
+ # source: https://huggingface.co/TheBloke/stable-vicuna-13B-HF/discussions/5
1168
+ # source: https://huggingface.co/spaces/CarperAI/StableVicuna/blob/main/app.py
1169
+ register_conv_template(
1170
+ Conversation(
1171
+ name="stable-vicuna",
1172
+ system_message="### Assistant: I am StableVicuna, a large language model created by CarperAI. I am here to chat!\n",
1173
+ roles=("### Human", "### Assistant"),
1174
+ sep_style=SeparatorStyle.ADD_COLON_TWO,
1175
+ sep="\n",
1176
+ sep2="\n\n",
1177
+ )
1178
+ )
1179
+
1180
+ register_conv_template(
1181
+ Conversation(
1182
+ name="vigogne_chat_v3",
1183
+ system_template="[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n",
1184
+ system_message=(
1185
+ "Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez"
1186
+ " autant que vous le pouvez."
1187
+ ),
1188
+ roles=("[INST]", "[/INST]"),
1189
+ sep_style=SeparatorStyle.LLAMA2,
1190
+ sep=" ",
1191
+ sep2=" </s>",
1192
+ )
1193
+ )
1194
+
1195
+ # Falcon 180B chat template
1196
+ # source: https://huggingface.co/spaces/tiiuae/falcon-180b-demo/blob/d1590ee7fae9b6ce331ba7808e61a29dcce9239f/app.py#L28-L37
1197
+ register_conv_template(
1198
+ Conversation(
1199
+ name="falcon-chat",
1200
+ roles=("User", "Falcon"),
1201
+ system_template="System: {system_message}",
1202
+ messages=[],
1203
+ sep_style=SeparatorStyle.FALCON_CHAT,
1204
+ sep="\n",
1205
+ sep2="<|endoftext|>",
1206
+ stop_str="\nUser:", # use stop_str to stop generation after stop_token_ids, it will also remove stop_str from the generated text
1207
+ )
1208
+ )
1209
+
1210
+ # Phind template
1211
+ # source: https://huggingface.co/Phind/Phind-CodeLlama-34B-v2
1212
+ register_conv_template(
1213
+ Conversation(
1214
+ name="phind",
1215
+ system_message="### System Prompt\nYou are an intelligent programming assistant.",
1216
+ roles=("### User Message", "### Assistant"),
1217
+ messages=(),
1218
+ offset=0,
1219
+ sep_style=SeparatorStyle.ADD_COLON_SINGLE,
1220
+ sep="\n\n",
1221
+ )
1222
+ )
1223
+
1224
+ # Metharme formatting for Pygmalion models
1225
+ # source: https://huggingface.co/PygmalionAI/pygmalion-2-13b
1226
+ register_conv_template(
1227
+ Conversation(
1228
+ name="metharme",
1229
+ system_template="<|system|>{system_message}",
1230
+ system_message="""Enter RP mode. You shall reply to the user while staying
1231
+ in character. Your responses must be detailed, creative, immersive, and drive the scenario
1232
+ forward.""",
1233
+ roles=("<|user|>", "<|model|>"),
1234
+ sep_style=SeparatorStyle.NO_COLON_SINGLE,
1235
+ sep="",
1236
+ stop_str="<|user|>",
1237
+ )
1238
+ )
1239
+
1240
+ # Zephyr template
1241
+ # reference: https://huggingface.co/spaces/HuggingFaceH4/zephyr-playground/blob/main/dialogues.py
1242
+ register_conv_template(
1243
+ Conversation(
1244
+ name="zephyr",
1245
+ system_template="<|system|>\n{system_message}",
1246
+ roles=("<|user|>", "<|assistant|>"),
1247
+ sep_style=SeparatorStyle.CHATML,
1248
+ sep="</s>",
1249
+ stop_token_ids=[2],
1250
+ stop_str="</s>",
1251
+ )
1252
+ )
1253
+
1254
+ # Orca-2 template
1255
+ # reference: https://huggingface.co/microsoft/Orca-2-7b
1256
+ register_conv_template(
1257
+ Conversation(
1258
+ name="orca-2",
1259
+ system_template="<|im_start|>system\n{system_message}",
1260
+ system_message="You are Orca, an AI language model created by Microsoft. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.",
1261
+ roles=("<|im_start|>user", "<|im_start|>assistant"),
1262
+ sep_style=SeparatorStyle.CHATML,
1263
+ sep="<|im_end|>",
1264
+ stop_str="<|im_end|>",
1265
+ )
1266
+ )
1267
+
1268
+ if __name__ == "__main__":
1269
+ from fastchat.conversation import get_conv_template
1270
+
1271
+ print("-- Vicuna template --")
1272
+ conv = get_conv_template("vicuna_v1.1")
1273
+ conv.append_message(conv.roles[0], "Hello!")
1274
+ conv.append_message(conv.roles[1], "Hi!")
1275
+ conv.append_message(conv.roles[0], "How are you?")
1276
+ conv.append_message(conv.roles[1], None)
1277
+ print(conv.get_prompt())
1278
+
1279
+ print("\n")
1280
+
1281
+ print("-- Llama-2 template --")
1282
+ conv = get_conv_template("llama-2")
1283
+ conv.set_system_message("You are a helpful, respectful and honest assistant.")
1284
+ conv.append_message(conv.roles[0], "Hello!")
1285
+ conv.append_message(conv.roles[1], "Hi!")
1286
+ conv.append_message(conv.roles[0], "How are you?")
1287
+ conv.append_message(conv.roles[1], None)
1288
+ print(conv.get_prompt())
1289
+
1290
+ print("\n")
1291
+
1292
+ print("-- ChatGPT template --")
1293
+ conv = get_conv_template("chatgpt")
1294
+ conv.append_message(conv.roles[0], "Hello!")
1295
+ conv.append_message(conv.roles[1], "Hi!")
1296
+ conv.append_message(conv.roles[0], "How are you?")
1297
+ conv.append_message(conv.roles[1], None)
1298
+ print(conv.to_openai_api_messages())
1299
+
1300
+ print("\n")
1301
+
1302
+ print("-- Claude template --")
1303
+ conv = get_conv_template("claude")
1304
+ conv.append_message(conv.roles[0], "Hello!")
1305
+ conv.append_message(conv.roles[1], "Hi!")
1306
+ conv.append_message(conv.roles[0], "How are you?")
1307
+ conv.append_message(conv.roles[1], None)
1308
+ print(conv.get_prompt())
requirements-dev.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ fschat
requirements.txt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ transformers
2
+ streamlit
3
+ sentencepiece
templates/airoboros_v1.jinja2 ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {% if not loop.first %}
6
+ {% endif %}
7
+ {% if message['role'] == 'system' %}
8
+ {{ message['content'] + ' ' }}
9
+ {% elif message['role'] == 'user' %}
10
+ {{ 'USER: ' + message['content'] + ' ' }}
11
+ {% elif message['role'] == 'assistant' %}
12
+ {{ 'ASSISTANT: ' + message['content'] + '</s>' }}
13
+ {% endif %}
14
+ {% endfor %}
15
+ {% if add_generation_prompt %}
16
+ {{ 'ASSISTANT:' }}
17
+ {% endif %}
templates/airoboros_v2.jinja2 ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {% if not loop.first %}
6
+ {% endif %}
7
+ {% if message['role'] == 'system' %}
8
+ {{ message['content'] + '\n' }}
9
+ {% elif message['role'] == 'user' %}
10
+ {{ 'USER: ' + message['content'] + '\n' }}
11
+ {% elif message['role'] == 'assistant' %}
12
+ {{ 'ASSISTANT: ' + message['content'] + '</s>' }}
13
+ {% endif %}
14
+ {% endfor %}
15
+ {% if add_generation_prompt %}
16
+ {{ 'ASSISTANT:' }}
17
+ {% endif %}
templates/chat-ml.jinja2 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ {% for message in messages %}
2
+ {{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
3
+ {% endfor %}
templates/falcon-chat.jinja2 ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% for message in messages %}
2
+ {% if not loop.first %}
3
+ {{ '\n' }}
4
+ {% endif %}
5
+ {% if message['role'] == 'system' %}
6
+ {{ 'System: ' }}
7
+ {% elif message['role'] == 'user' %}
8
+ {{ 'User: ' }}
9
+ {% elif message['role'] == 'assistant' %}
10
+ {{ 'Falcon: ' }}
11
+ {% endif %}
12
+ {{ message['content'] }}
13
+ {% endfor %}
14
+ {% if add_generation_prompt %}
15
+ {{ '\n' + 'Falcon:' }}
16
+ {% endif %}
templates/llama-2.jinja2 ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {% if message['role'] == 'user' %}
6
+ {{ '<s>' + '[INST] ' + message['content'] + ' [/INST]' }}
7
+ {% elif message['role'] == 'system' %}
8
+ {{ '<<SYS>>\n' + message['content'] + '\n<</SYS>>\n\n' }}
9
+ {% elif message['role'] == 'assistant' %}
10
+ {{ ' ' + message['content'] + ' ' + '</s>' }}
11
+ {% endif %}
12
+ {% endfor %}
templates/mistral-7b-openorca.jinja2 ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
6
+ {% endfor %}
7
+ {% if add_generation_prompt %}
8
+ {{ '<|im_start|>assistant\n' }}
9
+ {% endif %}
templates/open-orca.jinja2 ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {% if message['role'] == 'user' %}
6
+ {{ 'User: ' + message['content'] + '<|end_of_turn|>' + '\n' }}
7
+ {% elif message['role'] == 'system' %}
8
+ {{ message['content'] + '<|end_of_turn|>' + '\n' }}
9
+ {% elif message['role'] == 'assistant' %}
10
+ {{ 'Assistant: ' + message['content'] + '<|end_of_turn|>' + '\n' }}
11
+ {% endif %}
12
+ {% endfor %}
13
+ {% if add_generation_prompt %}
14
+ {{ 'Assistant: ' }}
15
+ {% endif %}
templates/openhermes-2.5-mistral.jinja2 ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
6
+ {% endfor %}
7
+ {% if add_generation_prompt %}
8
+ {{ '<|im_start|>assistant\n' }}
9
+ {% endif %}
templates/orca2.jinja2 ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {{ '<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
6
+ {% endfor %}
7
+ {% if add_generation_prompt %}
8
+ {{ '<|im_start|>assistant\n' }}
9
+ {% endif %}
templates/vicuna_v1.1.jinja2 ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% if not add_generation_prompt is defined %}
2
+ {% set add_generation_prompt = false %}
3
+ {% endif %}
4
+ {% for message in messages %}
5
+ {% if not loop.first %}
6
+ {% endif %}
7
+ {% if message['role'] == 'system' %}
8
+ {{ message['content'] + ' ' }}
9
+ {% elif message['role'] == 'user' %}
10
+ {{ 'USER: ' + message['content'] + ' ' }}
11
+ {% elif message['role'] == 'assistant' %}
12
+ {{ 'ASSISTANT: ' + message['content'] + '</s>' }}
13
+ {% endif %}
14
+ {% endfor %}
15
+ {% if add_generation_prompt %}
16
+ {{ 'ASSISTANT:' }}
17
+ {% endif %}
templates/zephyr.jinja2 ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% for message in messages %}
2
+ {% if message['role'] == 'user' %}
3
+ {{ '<|user|>\n' + message['content'] + eos_token }}
4
+ {% elif message['role'] == 'system' %}
5
+ {{ '<|system|>\n' + message['content'] + eos_token }}
6
+ {% elif message['role'] == 'assistant' %}
7
+ {{ '<|assistant|>\n' + message['content'] + eos_token }}
8
+ {% endif %}
9
+ {% if loop.last and add_generation_prompt %}
10
+ {{ '<|assistant|>' }}
11
+ {% endif %}
12
+ {% endfor %}
tests_template/__pycache__/test_llama2.cpython-310-pytest-7.4.3.pyc ADDED
Binary file (2.13 kB). View file
 
tests_template/__pycache__/test_orca2 copy.cpython-310-pytest-7.4.3.pyc ADDED
Binary file (2.31 kB). View file
 
tests_template/__pycache__/test_orca2.cpython-310-pytest-7.4.3.pyc ADDED
Binary file (2.12 kB). View file
 
tests_template/__pycache__/utils.cpython-310.pyc ADDED
Binary file (375 Bytes). View file
 
tests_template/test_airoboros_v1.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+ import difflib
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/airoboros_v1.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="jondurbin/airoboros-l2-7b-gpt4-2.0", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ # tokenizer.eos_token = "<|end_of_turn|>"
30
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
31
+
32
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
33
+ print()
34
+ print("add_generation_prompt False:")
35
+ print(transformer_prompt)
36
+
37
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
38
+ print()
39
+ print("add_generation_prompt True:")
40
+ print(transformer_prompt)
41
+ # transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=True, add_generation_prompt=True)
42
+ # print(transformer_prompt)
43
+
44
+
45
+ print("Fastchat template: ")
46
+ conv = get_conv_template("airoboros_v1")
47
+
48
+ conv.set_system_message(chat[0]["content"])
49
+ conv.append_message(conv.roles[0], chat[1]["content"])
50
+ conv.append_message(conv.roles[1], chat[2]["content"])
51
+ conv.append_message(conv.roles[0], chat[3]["content"])
52
+ conv.append_message(conv.roles[1], None)
53
+ print(conv.get_prompt())
54
+ matcher = difflib.SequenceMatcher(a=transformer_prompt, b=conv.get_prompt())
55
+ print("Matching Sequences:")
56
+ for match in matcher.get_matching_blocks():
57
+ print("Match : {}".format(match))
58
+ print("Matching Sequence : {}".format(transformer_prompt[match.a:match.a+match.size]))
59
+ assert transformer_prompt == conv.get_prompt()
60
+
61
+ if __name__ == "__main__":
62
+ test_llama2_template()
tests_template/test_airoboros_v2.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+ import difflib
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/airoboros_v2.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="jondurbin/airoboros-l2-7b-2.2", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ # tokenizer.eos_token = "<|end_of_turn|>"
30
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
31
+
32
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
33
+ print()
34
+ print("add_generation_prompt False:")
35
+ print(transformer_prompt)
36
+
37
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
38
+ print()
39
+ print("add_generation_prompt True:")
40
+ print(transformer_prompt)
41
+ # transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=True, add_generation_prompt=True)
42
+ # print(transformer_prompt)
43
+
44
+
45
+ print("Fastchat template: ")
46
+ conv = get_conv_template("airoboros_v2")
47
+
48
+ conv.set_system_message(chat[0]["content"])
49
+ conv.append_message(conv.roles[0], chat[1]["content"])
50
+ conv.append_message(conv.roles[1], chat[2]["content"])
51
+ conv.append_message(conv.roles[0], chat[3]["content"])
52
+ conv.append_message(conv.roles[1], None)
53
+ print(conv.get_prompt())
54
+ matcher = difflib.SequenceMatcher(a=transformer_prompt, b=conv.get_prompt())
55
+ print("Matching Sequences:")
56
+ for match in matcher.get_matching_blocks():
57
+ print("Match : {}".format(match))
58
+ print("Matching Sequence : {}".format(transformer_prompt[match.a:match.a+match.size]))
59
+ assert transformer_prompt == conv.get_prompt()
60
+
61
+ if __name__ == "__main__":
62
+ test_llama2_template()
tests_template/test_llama2.py ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/llama-2.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="microsoft/Orca-2-7b", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ tokenizer.bos_token = "<s>"
30
+ tokenizer.eos_token = "</s>"
31
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
32
+
33
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
34
+ print()
35
+ print("add_generation_prompt False:")
36
+ print(transformer_prompt)
37
+
38
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
39
+ print()
40
+ print("add_generation_prompt True:")
41
+ print(transformer_prompt)
42
+
43
+
44
+ print("Fastchat template: ")
45
+ conv = get_conv_template("llama-2")
46
+
47
+ conv.set_system_message(chat[0]["content"])
48
+ conv.append_message(conv.roles[0], chat[1]["content"])
49
+ conv.append_message(conv.roles[1], chat[2]["content"])
50
+ conv.append_message(conv.roles[0], chat[3]["content"])
51
+ conv.append_message(conv.roles[1], None)
52
+ print(conv.get_prompt())
53
+
54
+ assert transformer_prompt == conv.get_prompt()
55
+
56
+ if __name__ == "__main__":
57
+ test_llama2_template()
tests_template/test_mistral-7b-openorca.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+ import difflib
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/mistral-7b-openorca.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="Open-Orca/Mistral-7B-OpenOrca", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ # tokenizer.eos_token = "<|end_of_turn|>"
30
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
31
+
32
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
33
+ print()
34
+ print("add_generation_prompt False:")
35
+ print(transformer_prompt)
36
+
37
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
38
+ print()
39
+ print("add_generation_prompt True:")
40
+ print(transformer_prompt)
41
+ # transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=True, add_generation_prompt=True)
42
+ # print(transformer_prompt)
43
+
44
+
45
+ print("Fastchat template: ")
46
+ conv = get_conv_template("mistral-7b-openorca")
47
+
48
+ conv.set_system_message(chat[0]["content"])
49
+ conv.append_message(conv.roles[0], chat[1]["content"])
50
+ conv.append_message(conv.roles[1], chat[2]["content"])
51
+ conv.append_message(conv.roles[0], chat[3]["content"])
52
+ conv.append_message(conv.roles[1], None)
53
+ print(conv.get_prompt())
54
+ matcher = difflib.SequenceMatcher(a=transformer_prompt, b=conv.get_prompt())
55
+ print("Matching Sequences:")
56
+ for match in matcher.get_matching_blocks():
57
+ print("Match : {}".format(match))
58
+ print("Matching Sequence : {}".format(transformer_prompt[match.a:match.a+match.size]))
59
+ assert transformer_prompt == conv.get_prompt()
60
+
61
+ if __name__ == "__main__":
62
+ test_llama2_template()
tests_template/test_openhermes-2.5-mistral.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+ import difflib
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/openhermes-2.5-mistral.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="teknium/OpenHermes-2.5-Mistral-7B", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ # tokenizer.eos_token = "<|end_of_turn|>"
30
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
31
+
32
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
33
+ print()
34
+ print("add_generation_prompt False:")
35
+ print(transformer_prompt)
36
+
37
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
38
+ print()
39
+ print("add_generation_prompt True:")
40
+ print(transformer_prompt)
41
+ # transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=True, add_generation_prompt=True)
42
+ # print(transformer_prompt)
43
+
44
+
45
+ print("Fastchat template: ")
46
+ conv = get_conv_template("OpenHermes-2.5-Mistral-7B")
47
+
48
+ conv.set_system_message(chat[0]["content"])
49
+ conv.append_message(conv.roles[0], chat[1]["content"])
50
+ conv.append_message(conv.roles[1], chat[2]["content"])
51
+ conv.append_message(conv.roles[0], chat[3]["content"])
52
+ conv.append_message(conv.roles[1], None)
53
+ print(conv.get_prompt())
54
+ matcher = difflib.SequenceMatcher(a=transformer_prompt, b=conv.get_prompt())
55
+ print("Matching Sequences:")
56
+ for match in matcher.get_matching_blocks():
57
+ print("Match : {}".format(match))
58
+ print("Matching Sequence : {}".format(transformer_prompt[match.a:match.a+match.size]))
59
+ assert transformer_prompt == conv.get_prompt()
60
+
61
+ if __name__ == "__main__":
62
+ test_llama2_template()
tests_template/test_openorca.py ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/open-orca.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="Open-Orca/OpenOrcaxOpenChat-Preview2-13B", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ tokenizer.eos_token = "<|end_of_turn|>"
30
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
31
+
32
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
33
+ print()
34
+ print("add_generation_prompt False:")
35
+ print(transformer_prompt)
36
+
37
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
38
+ print()
39
+ print("add_generation_prompt True:")
40
+ print(transformer_prompt)
41
+
42
+
43
+ print("Fastchat template: ")
44
+ conv = get_conv_template("open-orca")
45
+
46
+ conv.set_system_message(chat[0]["content"])
47
+ conv.append_message(conv.roles[0], chat[1]["content"])
48
+ conv.append_message(conv.roles[1], chat[2]["content"])
49
+ conv.append_message(conv.roles[0], chat[3]["content"])
50
+ conv.append_message(conv.roles[1], None)
51
+ print(conv.get_prompt())
52
+
53
+ assert transformer_prompt == conv.get_prompt()
54
+
55
+ if __name__ == "__main__":
56
+ test_llama2_template()
tests_template/test_orca2.py ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+
6
+ def test_orca2_template():
7
+ jinja_lines = []
8
+ with open("../templates/orca2.jinja2", "r") as f:
9
+ jinja_lines = f.readlines()
10
+
11
+ print("jinja_lines: ", jinja_lines)
12
+
13
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
14
+
15
+ chat = [
16
+ {"role": "system", "content": "You are a helpful assistant."},
17
+ {"role": "user", "content": "Hello, how are you?"},
18
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
19
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
20
+ ]
21
+
22
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="microsoft/Orca-2-7b", trust_remote_code=True)
23
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
24
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
25
+ # print(transformer_prompt)
26
+ # print(tokenizer.chat_template)
27
+
28
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
29
+
30
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
31
+ print("add_generation_prompt False:")
32
+ print(transformer_prompt)
33
+
34
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
35
+ print("add_generation_prompt True:")
36
+ print(transformer_prompt)
37
+
38
+
39
+ conv = get_conv_template("orca-2")
40
+ print("Fastchat template: ")
41
+
42
+ conv.set_system_message(chat[0]["content"])
43
+ conv.append_message(conv.roles[0], chat[1]["content"])
44
+ conv.append_message(conv.roles[1], chat[2]["content"])
45
+ conv.append_message(conv.roles[0], chat[3]["content"])
46
+ conv.append_message(conv.roles[1], None)
47
+ print(conv.get_prompt())
48
+
49
+ assert transformer_prompt == conv.get_prompt()
50
+
tests_template/test_vicuna_v1.1.py ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers import AutoTokenizer
2
+ from fastchat.conversation import get_conv_template
3
+ import os
4
+ from utils import sanitize_jinja2
5
+ import difflib
6
+
7
+ def test_llama2_template():
8
+ jinja_lines = []
9
+ with open("../templates/vicuna_v1.1.jinja2", "r") as f:
10
+ jinja_lines = f.readlines()
11
+
12
+ print("jinja_lines: ", jinja_lines)
13
+
14
+ print("sanitized: ", sanitize_jinja2(jinja_lines))
15
+
16
+ chat = [
17
+ {"role": "system", "content": "You are a helpful assistant."},
18
+ {"role": "user", "content": "Hello, how are you?"},
19
+ {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
20
+ {"role": "user", "content": "I'd like to show off how chat templating works!"},
21
+ ]
22
+
23
+ tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path="lmsys/vicuna-13b-v1.1", trust_remote_code=True)
24
+ # f"<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant"
25
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
26
+ print("default template")
27
+ print(transformer_prompt)
28
+ # print(tokenizer.chat_template)
29
+ # tokenizer.eos_token = "<|end_of_turn|>"
30
+ tokenizer.chat_template = sanitize_jinja2(jinja_lines)
31
+
32
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False)
33
+ print()
34
+ print("add_generation_prompt False:")
35
+ print(transformer_prompt)
36
+
37
+ transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
38
+ print()
39
+ print("add_generation_prompt True:")
40
+ print(transformer_prompt)
41
+ # transformer_prompt = tokenizer.apply_chat_template(chat, tokenize=True, add_generation_prompt=True)
42
+ # print(transformer_prompt)
43
+
44
+
45
+ print("Fastchat template: ")
46
+ conv = get_conv_template("vicuna_v1.1")
47
+
48
+ conv.set_system_message(chat[0]["content"])
49
+ conv.append_message(conv.roles[0], chat[1]["content"])
50
+ conv.append_message(conv.roles[1], chat[2]["content"])
51
+ conv.append_message(conv.roles[0], chat[3]["content"])
52
+ conv.append_message(conv.roles[1], None)
53
+ print(conv.get_prompt())
54
+ matcher = difflib.SequenceMatcher(a=transformer_prompt, b=conv.get_prompt())
55
+ print("Matching Sequences:")
56
+ for match in matcher.get_matching_blocks():
57
+ print("Match : {}".format(match))
58
+ print("Matching Sequence : {}".format(transformer_prompt[match.a:match.a+match.size]))
59
+ assert transformer_prompt == conv.get_prompt()
60
+
61
+ if __name__ == "__main__":
62
+ test_llama2_template()
tests_template/utils.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ def sanitize_jinja2(jinja_lines):
2
+
3
+ one_liner_jinja = ""
4
+ for line in jinja_lines:
5
+ one_liner_jinja += line.lstrip(" ").rstrip("\n")
6
+
7
+ return one_liner_jinja