Reminder
Description
In the coming v1, we plan to use hf official chat_template to manage utterances.
Make a proposal following.
It is easy for qwen3.6 template which will preserve history cot and won't change the history context.
Made cur_input_ids - pre_input_ids possible.
encode
import copy
import json
from datasets import load_dataset
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-35B-A3B")
v1_tool_dataset = load_dataset("llamafactory/reason-tool-use-demo-1500", split="train").select(range(8))
TOOLS = json.loads(v1_tool_dataset[0]["tools"])
v1_messages = v1_tool_dataset[0]["messages"]
def convert_to_hf_format(messages: list[dict]) -> list[dict]:
# v1 type -> hf message type
messages = copy.deepcopy(messages)
for i in range(len(messages)):
msg = messages[i]
contents = msg.get("content") or []
to_remove: list[int] = []
reasoning_chunks: list[str] = []
tool_calls: list[dict] = []
for j, content_item in enumerate(contents):
ctype = content_item.get("type")
if ctype in ("reasoning", "reasoning_content"):
reasoning_chunks.append(content_item["value"])
to_remove.append(j)
elif ctype == "tool_call":
tc = json.loads(content_item["value"])
tool_calls.append(
{"function": {"name": tc["name"], "arguments": tc.get("arguments", {})}}
)
to_remove.append(j)
elif ctype == "text":
content_item["text"] = content_item.pop("value", content_item.get("text", ""))
for j in reversed(to_remove):
del contents[j]
if len(contents) == 0:
del msg["content"]
if reasoning_chunks:
msg["reasoning_content"] = "\n".join(reasoning_chunks)
if tool_calls:
msg["tool_calls"] = tool_calls
return messages
def _encode(messages: list[dict]):
labels, input_ids = [], []
label_turns: list[list[int]] = []
n = len(messages)
kwargs = {
"tokenize": True,
"add_generation_prompt": False,
"tools": TOOLS,
"preserve_thinking": True,
"enable_thinking": True,
}
prefix_index = 0
# find the first query message and fill the prefix ids
for i in range(n):
if messages[i]["role"] == "user":
prefix_index = i
break
prefix_messages = messages[: prefix_index + 1]
prefix_input_ids = tokenizer.apply_chat_template(prefix_messages, **kwargs)["input_ids"]
input_ids.extend(prefix_input_ids)
labels.extend([-100] * len(input_ids))
prev_input_ids = prefix_input_ids
for i in range(prefix_index + 1, n):
cur_input_ids = tokenizer.apply_chat_template(messages[: i + 1], **kwargs)["input_ids"]
this_turn_len = len(cur_input_ids) - len(prev_input_ids)
turn_ids = cur_input_ids[-this_turn_len:]
input_ids.extend(turn_ids)
if messages[i]["role"] in ["user", "tool"]:
labels.extend([-100] * this_turn_len)
else:
labels.extend(turn_ids)
label_turns.append(turn_ids)
prev_input_ids = cur_input_ids
return labels, input_ids, label_turns
if __name__ == "__main__":
v1_messages = convert_to_hf_format(v1_messages)
labels, input_ids, label_turns = _encode(v1_messages)
# visualize the input_ids and labels
print("###input_ids:")
print(tokenizer.decode(input_ids))
print("=" * 100)
print(f"####labels: ({len(label_turns)} supervised turn(s))")
for idx, turn_ids in enumerate(label_turns):
print(f"\n----- turn {idx + 1} / {len(label_turns)} (len={len(turn_ids)}) -----")
print(tokenizer.decode(turn_ids))
assert len(labels) == len(input_ids)output
###input_ids:
<|im_start|>system
# Tools
You have access to the following functions:
<tools>
{"type": "function", "function": {"name": "calculate_factorial", "description": "Calculates the factorial of a non-negative integer.\n\n Args:\n n (int): The non-negative integer.\n\n Returns:\n int: The factorial of n.\n\n Raises:\n ValueError: If n is negative.", "parameters": {"type": "object", "properties": {"n": {"type": "integer", "description": "The non-negative integer."}}, "required": ["n"]}}}
</tools>
If you choose to call a function ONLY reply in the following format with NO suffix:
<tool_call>
<function=example_function_name>
<parameter=example_parameter_1>
value_1
</parameter>
<parameter=example_parameter_2>
This is the value for the second parameter
that can span
multiple lines
</parameter>
</function>
</tool_call>
<IMPORTANT>
Reminder:
- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags
- Required parameters MUST be specified
- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after
- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls
</IMPORTANT>
You are a methodical and expert assistant. Your primary goal is to solve user requests by leveraging a set of available tools. You must reason for the best course of action in a structured manner before responding.<|im_end|>
<|im_start|>user
Calculate the factorial for 25.<|im_end|>
<|im_start|>assistant
<think>
Alright, the user wants me to calculate the factorial of 25. Since there's a 'calculate_factorial' tool available, I can use that. Factorial of 25 is a large number, but the tool should handle it. Let me just call the function with n=25.
</think>
<tool_call>
<function=calculate_factorial>
<parameter=n>
25
</parameter>
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
[{"tool_call_id": "call_C5l3iwmhActJBZnZsg4PTymF", "name": "calculate_factorial", "content": "{'factorial': 15511210043330985984000000, 'n': 25}"}]
</tool_response><|im_end|>
<|im_start|>assistant
<think>
The assistant was given a factorial calculation for 25. The function calculate_factorial was called with n=25. The result is a very large number, 15511210043330985984000000. There's no need for further processing here, just present the result as is.
</think>
The factorial of 25 is 15511210043330985984000000.<|im_end|>
<|im_start|>user
I need to calculate the factorial for another number. Please determine the factorial of 30 using the provided function.<|im_end|>
<|im_start|>assistant
<think>
The user wants to compute the factorial of 30, which is a larger number than the previous request. I'll use the same 'calculate_factorial' function by providing n=30. Let's see what the result is.
</think>
<tool_call>
<function=calculate_factorial>
<parameter=n>
30
</parameter>
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
[{"tool_call_id": "call_Jd6idzuU1bTksV68EsijgP4m", "name": "calculate_factorial", "content": "{'factorial': 265252859812191058636308480000000, 'n': 30}"}]
</tool_response><|im_end|>
<|im_start|>assistant
<think>
The user has requested the factorial of 30. The calculation was performed using the 'calculate_factorial' function provided, and the result is 265252859812191058636308480000000. This is a very large number, typical for factorials of numbers in the 30s.
</think>
The factorial of 30 is 265252859812191058636308480000000.<|im_end|>
====================================================================================================
####labels: (4 supervised turn(s))
----- turn 1 / 4 (len=99) -----
<|im_start|>assistant
<think>
Alright, the user wants me to calculate the factorial of 25. Since there's a 'calculate_factorial' tool available, I can use that. Factorial of 25 is a large number, but the tool should handle it. Let me just call the function with n=25.
</think>
<tool_call>
<function=calculate_factorial>
<parameter=n>
25
</parameter>
</function>
</tool_call><|im_end|>
----- turn 2 / 4 (len=122) -----
<|im_start|>assistant
<think>
The assistant was given a factorial calculation for 25. The function calculate_factorial was called with n=25. The result is a very large number, 15511210043330985984000000. There's no need for further processing here, just present the result as is.
</think>
The factorial of 25 is 15511210043330985984000000.<|im_end|>
----- turn 3 / 4 (len=85) -----
<|im_start|>assistant
<think>
The user wants to compute the factorial of 30, which is a larger number than the previous request. I'll use the same 'calculate_factorial' function by providing n=30. Let's see what the result is.
</think>
<tool_call>
<function=calculate_factorial>
<parameter=n>
30
</parameter>
</function>
</tool_call><|im_end|>
----- turn 4 / 4 (len=136) -----
<|im_start|>assistant
<think>
The user has requested the factorial of 30. The calculation was performed using the 'calculate_factorial' function provided, and the result is 265252859812191058636308480000000. This is a very large number, typical for factorials of numbers in the 30s.
</think>
The factorial of 30 is 265252859812191058636308480000000.<|im_end|>
cc @hiyouga @jiaqiw09 for suggestion
Pull Request
No response
Reminder
Description
In the coming v1, we plan to use hf official chat_template to manage utterances.
Make a proposal following.
encode
output
cc @hiyouga @jiaqiw09 for suggestion
Pull Request
No response