73 lines
1.7 KiB
Python
73 lines
1.7 KiB
Python
from __future__ import annotations
|
|
|
|
import torch
|
|
from transformers import StoppingCriteria
|
|
|
|
|
|
def contains_complete_json_object(text: str) -> bool:
|
|
started = False
|
|
depth = 0
|
|
in_string = False
|
|
escaped = False
|
|
|
|
for character in text:
|
|
if not started:
|
|
if character == "{":
|
|
started = True
|
|
depth = 1
|
|
|
|
continue
|
|
|
|
if in_string:
|
|
if escaped:
|
|
escaped = False
|
|
elif character == "\\":
|
|
escaped = True
|
|
elif character == '"':
|
|
in_string = False
|
|
|
|
continue
|
|
|
|
if character == '"':
|
|
in_string = True
|
|
elif character == "{":
|
|
depth += 1
|
|
elif character == "}":
|
|
depth -= 1
|
|
|
|
if depth == 0:
|
|
return True
|
|
|
|
return False
|
|
|
|
|
|
class JsonObjectStoppingCriteria(StoppingCriteria):
|
|
def __init__(self, tokenizer, prompt_length: int) -> None:
|
|
self.tokenizer = tokenizer
|
|
self.prompt_length = prompt_length
|
|
|
|
def __call__(
|
|
self,
|
|
input_ids: torch.LongTensor,
|
|
scores: torch.FloatTensor | None,
|
|
**kwargs,
|
|
) -> torch.BoolTensor:
|
|
completed = []
|
|
|
|
for row in input_ids:
|
|
generated_text = self.tokenizer.decode(
|
|
row[self.prompt_length:],
|
|
skip_special_tokens=True,
|
|
clean_up_tokenization_spaces=False,
|
|
)
|
|
|
|
completed.append(
|
|
contains_complete_json_object(generated_text)
|
|
)
|
|
|
|
return torch.tensor(
|
|
completed,
|
|
dtype=torch.bool,
|
|
device=input_ids.device,
|
|
)
|