Create tokenization.py
Browse files- tokenization.py +53 -0
tokenization.py
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import PreTrainedTokenizerFast
|
| 2 |
+
|
| 3 |
+
class SinglePOTokenizer(PreTrainedTokenizerFast):
|
| 4 |
+
def __init__(self, *args, **kwargs):
|
| 5 |
+
super().__init__(*args, **kwargs)
|
| 6 |
+
|
| 7 |
+
def get_context(
|
| 8 |
+
self,
|
| 9 |
+
raw_instruction,
|
| 10 |
+
rule_description,
|
| 11 |
+
):
|
| 12 |
+
prompt = "You are an expert prompt engineer." + " "
|
| 13 |
+
prompt += "Please help me optimize this prompt to get better response:\n\n[The Start of Raw Prompt]\n{}\n[The End of Raw Prompt]".format(raw_instruction)
|
| 14 |
+
prompt += "\n\nYou should optimize this prompt by {}".format(rule_description)
|
| 15 |
+
|
| 16 |
+
context = self.apply_chat_template(
|
| 17 |
+
[
|
| 18 |
+
{
|
| 19 |
+
"role": "user",
|
| 20 |
+
"content": prompt,
|
| 21 |
+
}
|
| 22 |
+
],
|
| 23 |
+
add_generation_prompt=True,
|
| 24 |
+
tokenize=False,
|
| 25 |
+
) + "The Optimized Prompt:\n\n[The Start of Optimized Prompt"
|
| 26 |
+
|
| 27 |
+
return context
|
| 28 |
+
|
| 29 |
+
def parse_output(
|
| 30 |
+
self,
|
| 31 |
+
output_text,
|
| 32 |
+
raw_instruction = "", # recommend to provide, so when some error happened, we can still use the raw instruction
|
| 33 |
+
):
|
| 34 |
+
better_instruction = "The Optimized Prompt:\n\n[The Start of Optimized Prompt" + output_text
|
| 35 |
+
|
| 36 |
+
if "[The Start of Optimized Prompt]" in better_instruction:
|
| 37 |
+
better_instruction = better_instruction[better_instruction.index("[The Start of Optimized Prompt]") + len("[The Start of Optimized Prompt]"):]
|
| 38 |
+
if better_instruction.startswith("\n"):
|
| 39 |
+
better_instruction = better_instruction[1:]
|
| 40 |
+
if "[The End of Optimized Prompt]" in better_instruction:
|
| 41 |
+
better_instruction = better_instruction[:better_instruction.index("[The End of Optimized Prompt]")]
|
| 42 |
+
if better_instruction.endswith("\n"):
|
| 43 |
+
better_instruction = better_instruction[:-1]
|
| 44 |
+
if "The Optimized Prompt:" in better_instruction: # almost error happened
|
| 45 |
+
better_instruction = better_instruction[:better_instruction.index("The Optimized Prompt:")]
|
| 46 |
+
|
| 47 |
+
if better_instruction.strip() == "": # some error may happen in optimization, so use the raw instruction
|
| 48 |
+
better_instruction = raw_instruction
|
| 49 |
+
|
| 50 |
+
if "The Optimized" in better_instruction: # still some error happened
|
| 51 |
+
better_instruction = raw_instruction
|
| 52 |
+
|
| 53 |
+
return better_instruction
|