jeffreymeetkai
commited on
Commit
•
29db9eb
1
Parent(s):
0fb49ab
Upload folder using huggingface_hub
Browse files- README.md +34 -13
- modeling_functionary.py +5 -4
- tokenization_functionary.py +57 -369
- tokenizer_config.json +1 -1
README.md
CHANGED
@@ -28,8 +28,8 @@ We provide custom code for both converting tool definitions into the system prom
|
|
28 |
```python
|
29 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
30 |
|
31 |
-
tokenizer = AutoTokenizer.from_pretrained("meetkai/functionary-small-
|
32 |
-
model = AutoModelForCausalLM.from_pretrained("meetkai/functionary-small-
|
33 |
|
34 |
tools = [
|
35 |
{
|
@@ -61,9 +61,9 @@ print(tokenizer.decode(pred.cpu()[0]))
|
|
61 |
|
62 |
## Prompt Template
|
63 |
|
64 |
-
We convert function definitions to a similar text to
|
65 |
|
66 |
-
This formatting is also available via our vLLM server which we process the functions into
|
67 |
|
68 |
```python
|
69 |
from openai import OpenAI
|
@@ -101,21 +101,42 @@ will yield:
|
|
101 |
```
|
102 |
<|start_header_id|>system<|end_header_id|>
|
103 |
|
104 |
-
|
105 |
-
namespace functions {
|
106 |
|
107 |
-
|
108 |
-
type get_current_weather = (_: {
|
109 |
-
// The city and state, e.g. San Francisco, CA
|
110 |
-
location: string,
|
111 |
-
}) => any;
|
112 |
|
113 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
What is the weather for Istanbul?
|
116 |
```
|
117 |
|
118 |
-
A more detailed example is provided [here](https://github.com/MeetKai/functionary/blob/main/tests/
|
119 |
|
120 |
## Run the model
|
121 |
|
|
|
28 |
```python
|
29 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
30 |
|
31 |
+
tokenizer = AutoTokenizer.from_pretrained("meetkai/functionary-small-v3.1", trust_remote_code=True)
|
32 |
+
model = AutoModelForCausalLM.from_pretrained("meetkai/functionary-small-v3.1", device_map="auto", trust_remote_code=True)
|
33 |
|
34 |
tools = [
|
35 |
{
|
|
|
61 |
|
62 |
## Prompt Template
|
63 |
|
64 |
+
We convert function definitions to a similar text to Meta's Llama 3.1 definitions. Then we inject these definitions as system prompts. After that, we inject the default system prompt. Then we start the conversation messages.
|
65 |
|
66 |
+
This formatting is also available via our vLLM server which we process the functions into definitions encapsulated in a system message and use a pre-defined Transformers chat template. This means that lists of messages can be formatted for you with the apply_chat_template() method within our server:
|
67 |
|
68 |
```python
|
69 |
from openai import OpenAI
|
|
|
101 |
```
|
102 |
<|start_header_id|>system<|end_header_id|>
|
103 |
|
104 |
+
Environment: ipython
|
|
|
105 |
|
106 |
+
Cutting Knowledge Date: December 2023
|
|
|
|
|
|
|
|
|
107 |
|
108 |
+
|
109 |
+
You have access to the following functions:
|
110 |
+
|
111 |
+
Use the function 'get_current_weather' to 'Get the current weather'
|
112 |
+
{"name": "get_current_weather", "description": "Get the current weather", "parameters": {"type": "object", "properties": {"location": {"type": "string", "description": "The city and state, e.g. San Francisco, CA"}},"required": ["location"]}}
|
113 |
+
|
114 |
+
|
115 |
+
Think very carefully before calling functions.
|
116 |
+
If a you choose to call a function ONLY reply in the following format:
|
117 |
+
<{start_tag}={function_name}>{parameters}{end_tag}
|
118 |
+
where
|
119 |
+
|
120 |
+
start_tag => `<function`
|
121 |
+
parameters => a JSON dict with the function argument name as key and function argument value as value.
|
122 |
+
end_tag => `</function>`
|
123 |
+
|
124 |
+
Here is an example,
|
125 |
+
<function=example_function_name>{"example_name": "example_value"}</function>
|
126 |
+
|
127 |
+
Reminder:
|
128 |
+
- If looking for real time information use relevant functions before falling back to brave_search
|
129 |
+
- Function calls MUST follow the specified format, start with <function= and end with </function>
|
130 |
+
- Required parameters MUST be specified
|
131 |
+
- Only call one function at a time
|
132 |
+
- Put the entire function call reply on one line
|
133 |
+
|
134 |
+
<|eot_id|><|start_header_id|>user<|end_header_id|>
|
135 |
|
136 |
What is the weather for Istanbul?
|
137 |
```
|
138 |
|
139 |
+
A more detailed example is provided [here](https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt).
|
140 |
|
141 |
## Run the model
|
142 |
|
modeling_functionary.py
CHANGED
@@ -66,7 +66,7 @@ class FunctionaryForCausalLM(LlamaForCausalLM):
|
|
66 |
)
|
67 |
|
68 |
input_ids = kwargs.pop("input_ids")
|
69 |
-
function_call_token = "
|
70 |
|
71 |
correct_results = []
|
72 |
for input_id, result in zip(input_ids, results):
|
@@ -80,13 +80,14 @@ class FunctionaryForCausalLM(LlamaForCausalLM):
|
|
80 |
continue
|
81 |
|
82 |
chunk = chunk.replace(tokenizer.pad_token, "")
|
|
|
83 |
if i == 0 and has_text is not False:
|
84 |
-
final_output_json["content"] = chunk.
|
85 |
else:
|
86 |
tool_calls.append(
|
87 |
{
|
88 |
-
"name": chunk[: chunk.index("
|
89 |
-
"arguments": chunk[chunk.index("
|
90 |
}
|
91 |
)
|
92 |
if len(tool_calls) > 0:
|
|
|
66 |
)
|
67 |
|
68 |
input_ids = kwargs.pop("input_ids")
|
69 |
+
function_call_token = "<function="
|
70 |
|
71 |
correct_results = []
|
72 |
for input_id, result in zip(input_ids, results):
|
|
|
80 |
continue
|
81 |
|
82 |
chunk = chunk.replace(tokenizer.pad_token, "")
|
83 |
+
|
84 |
if i == 0 and has_text is not False:
|
85 |
+
final_output_json["content"] = chunk.removesuffix("<|eom_id|>").removesuffix("<|eot_id|>")
|
86 |
else:
|
87 |
tool_calls.append(
|
88 |
{
|
89 |
+
"name": chunk[: chunk.index(">{")],
|
90 |
+
"arguments": chunk[chunk.index(">{") + 1: ].removesuffix("<|eom_id|>").removesuffix("</function>")
|
91 |
}
|
92 |
)
|
93 |
if len(tool_calls) > 0:
|
tokenization_functionary.py
CHANGED
@@ -1,6 +1,7 @@
|
|
1 |
# Copyright (c) 2024, MeetKai Inc. All rights reserved.
|
2 |
|
3 |
from copy import deepcopy
|
|
|
4 |
import json
|
5 |
from typing import Any, Dict, List, Literal, Optional, Union
|
6 |
|
@@ -14,382 +15,69 @@ from transformers.utils import TensorType, logging
|
|
14 |
|
15 |
|
16 |
logger = logging.get_logger(__name__)
|
17 |
-
SYSTEM_PROMPT = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. The assistant calls functions with appropriate input when necessary"""
|
18 |
-
CODE_INTERPRETER_SYSTEM_PROMPT = """When you send a message containing Python code to python, it will be executed in a stateful Jupyter notebook environment. python will respond with the output of the execution or time out after 60.0 seconds. The drive at '/mnt/data' can be used to save and persist user files."""
|
19 |
|
20 |
-
|
21 |
-
name
|
22 |
-
|
23 |
-
|
24 |
-
|
|
|
25 |
|
26 |
-
class Tool(BaseModel):
|
27 |
-
type: Literal["function", "code_interpreter"]
|
28 |
-
function: Optional[Function] = None
|
29 |
-
|
30 |
-
@model_validator(mode="after")
|
31 |
-
def check_type_function_matches(self) -> Self:
|
32 |
-
if self.type == "function":
|
33 |
-
assert self.function is not None, '"function" must contain function description when `"type": "function"`'
|
34 |
-
else:
|
35 |
-
assert self.function is None, '"function" must not be provided when `"type": "code_interpreter"`'
|
36 |
-
return self
|
37 |
-
|
38 |
-
|
39 |
-
def convert_data_type(param_type: str) -> str:
|
40 |
-
"""convert data_type to typescript data type
|
41 |
-
|
42 |
-
Args:
|
43 |
-
param_type (str): param_type
|
44 |
-
|
45 |
-
Returns:
|
46 |
-
str: param type in typescript
|
47 |
-
"""
|
48 |
-
if param_type == "integer" or param_type == "float":
|
49 |
-
return "number"
|
50 |
-
return param_type
|
51 |
-
|
52 |
-
|
53 |
-
def get_param_type(param: Dict) -> str:
|
54 |
-
"""get param_type of parameter
|
55 |
-
|
56 |
-
Args:
|
57 |
-
param (Dict): param dict in properties
|
58 |
-
|
59 |
-
Returns:
|
60 |
-
str: _description_
|
61 |
-
"""
|
62 |
-
param_type = "any"
|
63 |
-
if "type" in param:
|
64 |
-
raw_param_type = param["type"]
|
65 |
-
if type(raw_param_type) is list:
|
66 |
-
param_type = " | ".join(raw_param_type)
|
67 |
-
else:
|
68 |
-
param_type = raw_param_type
|
69 |
-
|
70 |
-
else: # in many cases, the json schema contains: oneOf instead of "type"
|
71 |
-
if "oneOf" in param:
|
72 |
-
one_of_types = []
|
73 |
-
for item in param["oneOf"]:
|
74 |
-
if "type" in item:
|
75 |
-
one_of_types.append(convert_data_type(item["type"]))
|
76 |
-
one_of_types = list(set(one_of_types))
|
77 |
-
param_type = " | ".join(one_of_types)
|
78 |
-
return convert_data_type(param_type)
|
79 |
-
|
80 |
-
|
81 |
-
def get_format_param(param: Dict) -> Optional[str]:
|
82 |
-
"""Get "format" from param. There are cases where format is not directly in param but in oneOf
|
83 |
-
|
84 |
-
Args:
|
85 |
-
param (Dict): _description_
|
86 |
-
|
87 |
-
Returns:
|
88 |
-
Optional[str]: _description_
|
89 |
-
"""
|
90 |
-
if "format" in param:
|
91 |
-
return param["format"]
|
92 |
-
if "oneOf" in param:
|
93 |
-
formats = []
|
94 |
-
for item in param["oneOf"]:
|
95 |
-
if "format" in item:
|
96 |
-
formats.append(item["format"])
|
97 |
-
if len(formats) > 0:
|
98 |
-
return " or ".join(formats)
|
99 |
-
return None
|
100 |
-
|
101 |
-
|
102 |
-
def get_param_info(param: Dict) -> Optional[str]:
|
103 |
-
"""get additional information about parameter such as: format, default value, min, max, ...
|
104 |
-
|
105 |
-
Args:
|
106 |
-
param (Dict): _description_
|
107 |
-
|
108 |
-
Returns:
|
109 |
-
Optional[str]: _description_
|
110 |
-
"""
|
111 |
-
param_type = param.get("type", "any")
|
112 |
-
info_list = []
|
113 |
-
if "description" in param:
|
114 |
-
desc = param["description"]
|
115 |
-
if not desc.endswith("."):
|
116 |
-
desc += "."
|
117 |
-
info_list.append(desc)
|
118 |
-
|
119 |
-
if "default" in param:
|
120 |
-
default_value = param["default"]
|
121 |
-
if param_type == "string":
|
122 |
-
default_value = f'"{default_value}"' # if string --> add ""
|
123 |
-
info_list.append(f"Default={default_value}.")
|
124 |
-
|
125 |
-
format_param = get_format_param(param)
|
126 |
-
if format_param is not None:
|
127 |
-
info_list.append("Format=" + format_param)
|
128 |
-
|
129 |
-
for field, field_name in [
|
130 |
-
("maximum", "Maximum"),
|
131 |
-
("minimum", "Minimum"),
|
132 |
-
("maxLength", "Maximum length"),
|
133 |
-
("minLength", "Minimum length"),
|
134 |
-
]:
|
135 |
-
if field in param:
|
136 |
-
info_list.append(f"{field_name}=" + str(param[field]))
|
137 |
-
|
138 |
-
if len(info_list) > 0:
|
139 |
-
result = "// " + " ".join(info_list)
|
140 |
-
result = result.replace("\n", " ")
|
141 |
-
return result
|
142 |
-
return None
|
143 |
-
|
144 |
-
|
145 |
-
def append_new_param_info(
|
146 |
-
info_list: List[str],
|
147 |
-
param_declaration: str,
|
148 |
-
comment_info: Optional[str],
|
149 |
-
examples_info: List,
|
150 |
-
depth: int,
|
151 |
-
):
|
152 |
-
"""Append a new parameter with comment to the info_list
|
153 |
-
|
154 |
-
Args:
|
155 |
-
info_lines (List[str]): current info_list
|
156 |
-
param_declaration (str): param: type
|
157 |
-
comment_info (Optional[str]): information of comment
|
158 |
-
examples_info (List): information of examples given
|
159 |
-
depth (int): level of nested param
|
160 |
-
"""
|
161 |
-
offset = ""
|
162 |
-
if depth >= 1:
|
163 |
-
offset = "".join([" " for _ in range(depth)])
|
164 |
-
if comment_info is not None:
|
165 |
-
# if depth == 0: # format: //comment\nparam: type
|
166 |
-
info_list.append(f"{offset}{comment_info}")
|
167 |
-
if len(examples_info) > 0:
|
168 |
-
for example in examples_info:
|
169 |
-
info_list.append(f"{offset}{example}")
|
170 |
-
info_list.append(f"{offset}{param_declaration}")
|
171 |
-
# else: # format: param: type // comment
|
172 |
-
# info_list.append(f"{offset}{param_declaration} {comment_info}")
|
173 |
-
else:
|
174 |
-
info_list.append(f"{offset}{param_declaration}")
|
175 |
-
|
176 |
-
|
177 |
-
def get_examples_info(param_name: str, examples: List) -> List:
|
178 |
-
"""get information about examples provided
|
179 |
-
|
180 |
-
Args:
|
181 |
-
param_name (str): _description_
|
182 |
-
examples (List): _description_
|
183 |
|
184 |
-
|
185 |
-
|
186 |
-
"""
|
187 |
-
examples_list = [f"// Example {param_name}:"]
|
188 |
-
for example in examples:
|
189 |
-
if isinstance(example, dict) or isinstance(example, list):
|
190 |
-
example_str = json.dumps(example, ensure_ascii=False).replace('\n', '\\n')
|
191 |
-
else:
|
192 |
-
example_str = str(example).replace('\n', '\\n')
|
193 |
-
examples_list.append(f"// {example_str}")
|
194 |
-
|
195 |
-
return examples_list
|
196 |
|
197 |
|
198 |
-
def
|
199 |
-
|
|
|
|
|
|
|
200 |
|
201 |
-
|
202 |
-
|
203 |
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
|
|
|
|
|
|
|
210 |
|
211 |
-
|
212 |
-
|
213 |
-
) -> str:
|
214 |
-
"""recursive implementation for generating type script of array
|
215 |
|
216 |
-
|
217 |
-
|
218 |
-
|
219 |
-
|
|
|
|
|
220 |
|
221 |
-
|
222 |
-
|
223 |
-
"""
|
224 |
-
offset = ""
|
225 |
-
if depth >= 1:
|
226 |
-
offset = "".join([" " for _ in range(depth)])
|
227 |
-
items_info = param_dic.get("items", {})
|
228 |
-
|
229 |
-
if len(items_info) == 0:
|
230 |
-
if param_name is not None:
|
231 |
-
return f"{offset}{param_name}: []"
|
232 |
-
else:
|
233 |
-
return "[]"
|
234 |
-
array_type = get_param_type(items_info)
|
235 |
-
if array_type == "object":
|
236 |
-
info_lines = []
|
237 |
-
child_lines = get_parameter_typescript(
|
238 |
-
items_info.get("properties", {}), items_info.get("required", []), depth + 1
|
239 |
-
)
|
240 |
-
# if comment_info is not None:
|
241 |
-
# info_lines.append(f"{offset}{comment_info}")
|
242 |
-
if param_name is not None:
|
243 |
-
info_lines.append(f"{offset}{param_name}" + ": {")
|
244 |
-
else:
|
245 |
-
info_lines.append(f"{offset}" + "{")
|
246 |
-
info_lines.extend(child_lines)
|
247 |
-
info_lines.append(f"{offset}" + "}[]")
|
248 |
-
return "\n".join(info_lines)
|
249 |
|
250 |
-
elif array_type == "array":
|
251 |
-
item_info = get_array_typescript(None, items_info, depth + 1)
|
252 |
-
if param_name is None:
|
253 |
-
return f"{item_info}[]"
|
254 |
-
return f"{offset}{param_name}: {item_info.strip()}[]"
|
255 |
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
return f"({item_type})[]"
|
261 |
-
else:
|
262 |
-
return f"{offset}{param_name}: ({item_type})[]"
|
263 |
-
else:
|
264 |
-
if param_name is None:
|
265 |
-
return f"{array_type}[]"
|
266 |
-
else:
|
267 |
-
return f"{offset}{param_name}: {array_type}[],"
|
268 |
|
|
|
|
|
|
|
|
|
|
|
269 |
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
|
274 |
-
|
275 |
-
properties (_type_): properties in parameters
|
276 |
-
required_params (_type_): List of required parameters
|
277 |
-
depth (int, optional): the depth of params (nested level). Defaults to 0.
|
278 |
|
279 |
-
Returns:
|
280 |
-
_type_: list of lines containing information about all parameters
|
281 |
-
"""
|
282 |
-
tp_lines = []
|
283 |
-
for param_name, param in properties.items():
|
284 |
-
# Sometimes properties have "required" field as a list of string.
|
285 |
-
# Even though its supposed to be not under properties. So we skip it
|
286 |
-
if not isinstance(param, dict):
|
287 |
-
continue
|
288 |
-
# Param Description
|
289 |
-
comment_info = get_param_info(param)
|
290 |
-
# Param Examples
|
291 |
-
examples_info = []
|
292 |
-
if "examples" in param:
|
293 |
-
examples_info = get_examples_info(param_name, param["examples"])
|
294 |
-
# Param Name declaration
|
295 |
-
param_declaration = f"{param_name}"
|
296 |
-
if isinstance(required_params, list):
|
297 |
-
if param_name not in required_params:
|
298 |
-
param_declaration += "?"
|
299 |
-
param_type = get_param_type(param)
|
300 |
-
|
301 |
-
offset = ""
|
302 |
-
if depth >= 1:
|
303 |
-
offset = "".join([" " for _ in range(depth)])
|
304 |
-
|
305 |
-
if param_type == "object": # param_type is object
|
306 |
-
child_lines = get_parameter_typescript(
|
307 |
-
param.get("properties", {}), param.get("required", []), depth + 1
|
308 |
-
)
|
309 |
-
if comment_info is not None:
|
310 |
-
tp_lines.append(f"{offset}{comment_info}")
|
311 |
-
if len(examples_info) > 0:
|
312 |
-
for example in examples_info:
|
313 |
-
tp_lines.append(f"{offset}{example}")
|
314 |
-
|
315 |
-
param_declaration += ": {"
|
316 |
-
tp_lines.append(f"{offset}{param_declaration}")
|
317 |
-
tp_lines.extend(child_lines)
|
318 |
-
tp_lines.append(f"{offset}" + "},")
|
319 |
-
|
320 |
-
elif param_type == "array": # param_type is an array
|
321 |
-
item_info = param.get("items", {})
|
322 |
-
if "type" not in item_info: # don't know type of array
|
323 |
-
param_declaration += ": [],"
|
324 |
-
append_new_param_info(
|
325 |
-
tp_lines, param_declaration, comment_info, examples_info, depth
|
326 |
-
)
|
327 |
-
else:
|
328 |
-
array_declaration = get_array_typescript(
|
329 |
-
param_declaration, param, depth
|
330 |
-
)
|
331 |
-
if not array_declaration.endswith(","):
|
332 |
-
array_declaration += ","
|
333 |
-
if comment_info is not None:
|
334 |
-
tp_lines.append(f"{offset}{comment_info}")
|
335 |
-
if len(examples_info) > 0:
|
336 |
-
for example in examples_info:
|
337 |
-
tp_lines.append(f"{offset}{example}")
|
338 |
-
tp_lines.append(array_declaration)
|
339 |
-
else:
|
340 |
-
if "enum" in param:
|
341 |
-
param_type = get_enum_option_str(param["enum"])
|
342 |
-
# param_type = " | ".join([f'"{v}"' for v in param["enum"]])
|
343 |
-
if "nullable" in param and param["nullable"] is True:
|
344 |
-
param_type += " | null"
|
345 |
-
param_declaration += f": {param_type},"
|
346 |
-
append_new_param_info(
|
347 |
-
tp_lines, param_declaration, comment_info, examples_info, depth
|
348 |
-
)
|
349 |
-
|
350 |
-
return tp_lines
|
351 |
-
|
352 |
-
def generate_schema_from_functions(
|
353 |
-
functions: List[Function], namespace="functions"
|
354 |
-
) -> str:
|
355 |
-
"""
|
356 |
-
Convert functions schema to a schema that language models can understand.
|
357 |
-
"""
|
358 |
-
|
359 |
-
schema = "// Supported function definitions that should be called when necessary.\n"
|
360 |
-
schema += f"namespace {namespace} {{\n\n"
|
361 |
-
|
362 |
-
for function in functions:
|
363 |
-
# Convert a Function object to dict, if necessary
|
364 |
-
if not isinstance(function, dict):
|
365 |
-
function = function.model_dump()
|
366 |
-
function_name = function.get("name", None)
|
367 |
-
if function_name is None:
|
368 |
-
continue
|
369 |
-
|
370 |
-
description = function.get("description", "")
|
371 |
-
schema += f"// {description}\n"
|
372 |
-
schema += f"type {function_name}"
|
373 |
-
|
374 |
-
parameters = function.get("parameters", None)
|
375 |
-
if parameters is not None and parameters.get("properties") is not None:
|
376 |
-
parameters = deepcopy(jsonref.JsonRef.replace_refs(parameters))
|
377 |
-
schema += " = (_: {\n"
|
378 |
-
required_params = parameters.get("required", [])
|
379 |
-
tp_lines = get_parameter_typescript(
|
380 |
-
parameters.get("properties"),
|
381 |
-
required_params,
|
382 |
-
0,
|
383 |
-
)
|
384 |
-
schema += "\n".join(tp_lines)
|
385 |
-
schema += "\n}) => any;\n\n"
|
386 |
-
else:
|
387 |
-
# Doesn't have any parameters
|
388 |
-
schema += " = () => any;\n\n"
|
389 |
-
|
390 |
-
schema += f"}} // namespace {namespace}"
|
391 |
-
|
392 |
-
return schema
|
393 |
|
394 |
class FunctionaryTokenizer(PreTrainedTokenizerFast):
|
395 |
def apply_chat_template(
|
@@ -465,16 +153,16 @@ class FunctionaryTokenizer(PreTrainedTokenizerFast):
|
|
465 |
# Prepare tools/functions into schema
|
466 |
functions_pydantic_to_render = []
|
467 |
has_code_interpreter = False
|
468 |
-
|
469 |
-
|
470 |
-
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
|
475 |
-
|
476 |
-
|
477 |
-
conversation.insert(
|
478 |
|
479 |
# Compilation function uses a cache to avoid recompiling the same template
|
480 |
compiled_template = self._compile_jinja_template(chat_template)
|
|
|
1 |
# Copyright (c) 2024, MeetKai Inc. All rights reserved.
|
2 |
|
3 |
from copy import deepcopy
|
4 |
+
import datetime
|
5 |
import json
|
6 |
from typing import Any, Dict, List, Literal, Optional, Union
|
7 |
|
|
|
15 |
|
16 |
|
17 |
logger = logging.get_logger(__name__)
|
|
|
|
|
18 |
|
19 |
+
def get_instruction_string(custom_tool_definition) -> str:
|
20 |
+
name, description = (
|
21 |
+
custom_tool_definition["name"],
|
22 |
+
custom_tool_definition["description"],
|
23 |
+
)
|
24 |
+
return f"Use the function '{name}' to '{description}'"
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
+
def get_parameters_string(custom_tool_definition) -> str:
|
28 |
+
return json.dumps(custom_tool_definition)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
|
30 |
|
31 |
+
def get_system_prompt_for_custom_tools(custom_tools: List) -> str:
|
32 |
+
custom_tool_params = ""
|
33 |
+
for t in custom_tools:
|
34 |
+
custom_tool_params += get_instruction_string(t) + "\n"
|
35 |
+
custom_tool_params += get_parameters_string(t) + "\n\n"
|
36 |
|
37 |
+
content = f"""
|
38 |
+
You have access to the following functions:
|
39 |
|
40 |
+
{custom_tool_params}
|
41 |
+
Think very carefully before calling functions.
|
42 |
+
If a you choose to call a function ONLY reply in the following format:
|
43 |
+
<{{start_tag}}={{function_name}}>{{parameters}}{{end_tag}}
|
44 |
+
where
|
45 |
|
46 |
+
start_tag => `<function`
|
47 |
+
parameters => a JSON dict with the function argument name as key and function argument value as value.
|
48 |
+
end_tag => `</function>`
|
49 |
|
50 |
+
Here is an example,
|
51 |
+
<function=example_function_name>{{"example_name": "example_value"}}</function>
|
|
|
|
|
52 |
|
53 |
+
Reminder:
|
54 |
+
- If looking for real time information use relevant functions before falling back to brave_search
|
55 |
+
- Function calls MUST follow the specified format, start with <function= and end with </function>
|
56 |
+
- Required parameters MUST be specified
|
57 |
+
- Only call one function at a time
|
58 |
+
- Put the entire function call reply on one line
|
59 |
|
60 |
+
"""
|
61 |
+
return content
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
|
|
|
|
|
|
|
|
|
|
63 |
|
64 |
+
def get_system_message_for_tools(tools: List[Dict], use_code_interpreter) -> List[Dict]:
|
65 |
+
content = ""
|
66 |
+
if use_code_interpreter:
|
67 |
+
content += "Environment: ipython\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
|
69 |
+
current_date = datetime.datetime.now()
|
70 |
+
formatted_date = current_date.strftime("%d %B %Y")
|
71 |
+
date_str = f"""
|
72 |
+
Cutting Knowledge Date: December 2023\n\n"""
|
73 |
+
content += date_str
|
74 |
|
75 |
+
if tools:
|
76 |
+
custom_message = get_system_prompt_for_custom_tools(tools)
|
77 |
+
content += custom_message
|
78 |
|
79 |
+
return {"role": "system", "content": content}
|
|
|
|
|
|
|
80 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
class FunctionaryTokenizer(PreTrainedTokenizerFast):
|
83 |
def apply_chat_template(
|
|
|
153 |
# Prepare tools/functions into schema
|
154 |
functions_pydantic_to_render = []
|
155 |
has_code_interpreter = False
|
156 |
+
if tools is not None:
|
157 |
+
for item in tools:
|
158 |
+
if "function" in item and item["function"] is not None:
|
159 |
+
functions_pydantic_to_render.append(item["function"])
|
160 |
+
elif "type" in item and item["type"] == "code_interpreter":
|
161 |
+
has_code_interpreter = True
|
162 |
+
else:
|
163 |
+
functions_pydantic_to_render.append(item)
|
164 |
+
tools_system_message = get_system_message_for_tools(functions_pydantic_to_render, has_code_interpreter)
|
165 |
+
conversation.insert(0, tools_system_message)
|
166 |
|
167 |
# Compilation function uses a cache to avoid recompiling the same template
|
168 |
compiled_template = self._compile_jinja_template(chat_template)
|
tokenizer_config.json
CHANGED
@@ -2050,7 +2050,7 @@
|
|
2050 |
}
|
2051 |
},
|
2052 |
"bos_token": "<|begin_of_text|>",
|
2053 |
-
"chat_template": "",
|
2054 |
"clean_up_tokenization_spaces": true,
|
2055 |
"eos_token": "<|eot_id|>",
|
2056 |
"legacy": true,
|
|
|
2050 |
}
|
2051 |
},
|
2052 |
"bos_token": "<|begin_of_text|>",
|
2053 |
+
"chat_template": "{% for message in messages %}\n{% if message['role'] == 'user' or message['role'] == 'system' %}\n{{ '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n' + message['content'] + '<|eot_id|>' }}{% elif message['role'] == 'tool' %}\n{{ '<|start_header_id|>ipython<|end_header_id|>\n\n' + message['content'] + '<|eot_id|>' }}{% else %}\n{{ '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'}}{% if message['content'] is not none %}\n{{ message['content'] }}{% endif %}\n{% if 'tool_calls' in message and message['tool_calls'] is not none %}\n{% for tool_call in message['tool_calls'] %}\n{{ '<function=' + tool_call['function']['name'] + '>' + tool_call['function']['arguments'] + '</function>' }}{% endfor %}\n{{ '<|eom_id|>' }}{% else %}\n{{ '<|eot_id|>' }}\n{% endif %}{% endif %}\n{% endfor %}\n{% if add_generation_prompt %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}{% endif %}",
|
2054 |
"clean_up_tokenization_spaces": true,
|
2055 |
"eos_token": "<|eot_id|>",
|
2056 |
"legacy": true,
|