Refactor duplicate code between Prompter and Pygmalion
Browse files
src/axolotl/prompt_strategies/pygmalion.py
CHANGED
@@ -5,7 +5,11 @@ import logging
|
|
5 |
from collections import defaultdict
|
6 |
from typing import Generator
|
7 |
|
8 |
-
from axolotl.prompt_tokenizers import
|
|
|
|
|
|
|
|
|
9 |
|
10 |
IGNORE_TOKEN_ID = -100
|
11 |
|
@@ -23,12 +27,7 @@ class PygmalionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
23 |
self.bot_prefix_token_ids = res["input_ids"]
|
24 |
|
25 |
def tokenize_prompt(self, prompt):
|
26 |
-
result =
|
27 |
-
"input_ids": [],
|
28 |
-
"attention_mask": [],
|
29 |
-
"labels": [],
|
30 |
-
}
|
31 |
-
current_len = 0
|
32 |
for _, part in enumerate(self.prompter.build_prompt(prompt["conversations"])):
|
33 |
role, message = part
|
34 |
if role == "system":
|
@@ -67,37 +66,15 @@ class PygmalionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
67 |
else:
|
68 |
logging.warning(f"unknown role in conversation: {role}")
|
69 |
res = defaultdict(lambda: [])
|
70 |
-
input_ids = res["input_ids"]
|
71 |
-
input_len = len(input_ids)
|
72 |
-
result["input_ids"][current_len : current_len + input_len] = input_ids
|
73 |
-
result["attention_mask"][current_len : current_len + input_len] = [
|
74 |
-
1 if x != self.tokenizer.pad_token_id else 0 for x in input_ids
|
75 |
-
]
|
76 |
-
result["labels"][current_len : current_len + input_len] = labels
|
77 |
-
current_len += input_len
|
78 |
-
return result
|
79 |
-
|
80 |
-
def _tokenize(self, prompt, add_eos_token=True, strip_bos_token=False):
|
81 |
-
result = self.tokenizer(
|
82 |
-
prompt,
|
83 |
-
truncation=True,
|
84 |
-
max_length=self.sequence_len,
|
85 |
-
padding=False,
|
86 |
-
return_tensors=None,
|
87 |
-
)
|
88 |
-
if (
|
89 |
-
result["input_ids"][-1] != self.tokenizer.eos_token_id
|
90 |
-
and len(result["input_ids"]) < self.sequence_len
|
91 |
-
and add_eos_token
|
92 |
-
):
|
93 |
-
result["input_ids"].append(self.tokenizer.eos_token_id)
|
94 |
-
result["attention_mask"].append(1)
|
95 |
-
|
96 |
-
if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
|
97 |
-
result["input_ids"] = result["input_ids"][1:]
|
98 |
-
result["attention_mask"] = result["attention_mask"][1:]
|
99 |
|
100 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
return result
|
102 |
|
103 |
|
|
|
5 |
from collections import defaultdict
|
6 |
from typing import Generator
|
7 |
|
8 |
+
from axolotl.prompt_tokenizers import (
|
9 |
+
PromptTokenizingStrategy,
|
10 |
+
parse_tokenized_to_result,
|
11 |
+
tokenize_prompt_default,
|
12 |
+
)
|
13 |
|
14 |
IGNORE_TOKEN_ID = -100
|
15 |
|
|
|
27 |
self.bot_prefix_token_ids = res["input_ids"]
|
28 |
|
29 |
def tokenize_prompt(self, prompt):
|
30 |
+
result, current_len = tokenize_prompt_default()
|
|
|
|
|
|
|
|
|
|
|
31 |
for _, part in enumerate(self.prompter.build_prompt(prompt["conversations"])):
|
32 |
role, message = part
|
33 |
if role == "system":
|
|
|
66 |
else:
|
67 |
logging.warning(f"unknown role in conversation: {role}")
|
68 |
res = defaultdict(lambda: [])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
|
70 |
+
# pylint: disable=duplicate-code
|
71 |
+
result, current_len = parse_tokenized_to_result(
|
72 |
+
result,
|
73 |
+
current_len,
|
74 |
+
res,
|
75 |
+
labels,
|
76 |
+
pad_token_id=self.tokenizer.pad_token_id,
|
77 |
+
)
|
78 |
return result
|
79 |
|
80 |
|
src/axolotl/prompt_tokenizers.py
CHANGED
@@ -4,7 +4,7 @@ import abc
|
|
4 |
import copy
|
5 |
import functools
|
6 |
import logging
|
7 |
-
from typing import Tuple
|
8 |
|
9 |
from transformers import PreTrainedTokenizer
|
10 |
|
@@ -58,6 +58,29 @@ class PromptTokenizingStrategy(abc.ABC):
|
|
58 |
return id_or_ids
|
59 |
return False
|
60 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
|
62 |
class InstructionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
63 |
"""
|
@@ -106,29 +129,6 @@ class InstructionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
106 |
)
|
107 |
)
|
108 |
|
109 |
-
def _tokenize(self, prompt, add_eos_token=True, strip_bos_token=False):
|
110 |
-
result = self.tokenizer(
|
111 |
-
prompt,
|
112 |
-
truncation=True,
|
113 |
-
max_length=self.sequence_len,
|
114 |
-
padding=False,
|
115 |
-
return_tensors=None,
|
116 |
-
)
|
117 |
-
if (
|
118 |
-
result["input_ids"][-1] != self.tokenizer.eos_token_id
|
119 |
-
and len(result["input_ids"]) < self.sequence_len
|
120 |
-
and add_eos_token
|
121 |
-
):
|
122 |
-
result["input_ids"].append(self.tokenizer.eos_token_id)
|
123 |
-
result["attention_mask"].append(1)
|
124 |
-
|
125 |
-
if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
|
126 |
-
result["input_ids"] = result["input_ids"][1:]
|
127 |
-
result["attention_mask"] = result["attention_mask"][1:]
|
128 |
-
|
129 |
-
result["labels"] = result["input_ids"].copy()
|
130 |
-
return result
|
131 |
-
|
132 |
|
133 |
class AlpacaPromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
|
134 |
"""
|
@@ -295,7 +295,7 @@ class ReflectionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
295 |
)
|
296 |
)
|
297 |
|
298 |
-
def _tokenize(self, prompt, add_eos_token=True):
|
299 |
result = self.tokenizer(
|
300 |
prompt,
|
301 |
truncation=True,
|
@@ -339,12 +339,7 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
339 |
return prompt["conversations"]
|
340 |
|
341 |
def tokenize_prompt(self, prompt):
|
342 |
-
result =
|
343 |
-
"input_ids": [],
|
344 |
-
"attention_mask": [],
|
345 |
-
"labels": [],
|
346 |
-
}
|
347 |
-
current_len = 0
|
348 |
user_token = self._get_user_token()
|
349 |
assistant_token = self._get_assistant_token()
|
350 |
try:
|
@@ -382,14 +377,15 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
382 |
)
|
383 |
# everything from this is masked out from the labels
|
384 |
labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
|
385 |
-
|
386 |
-
|
387 |
-
result
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
|
392 |
-
|
|
|
393 |
return result
|
394 |
except (KeyError, AssertionError, IndexError) as err:
|
395 |
raise InvalidDataException(str(err)) from err
|
@@ -416,3 +412,40 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
|
|
416 |
|
417 |
result["labels"] = result["input_ids"].copy()
|
418 |
return result
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
import copy
|
5 |
import functools
|
6 |
import logging
|
7 |
+
from typing import Dict, List, Tuple
|
8 |
|
9 |
from transformers import PreTrainedTokenizer
|
10 |
|
|
|
58 |
return id_or_ids
|
59 |
return False
|
60 |
|
61 |
+
def _tokenize(self, prompt: str, add_eos_token=True, strip_bos_token=False):
|
62 |
+
result = self.tokenizer(
|
63 |
+
prompt,
|
64 |
+
truncation=True,
|
65 |
+
max_length=self.sequence_len,
|
66 |
+
padding=False,
|
67 |
+
return_tensors=None,
|
68 |
+
)
|
69 |
+
if (
|
70 |
+
result["input_ids"][-1] != self.tokenizer.eos_token_id
|
71 |
+
and len(result["input_ids"]) < self.sequence_len
|
72 |
+
and add_eos_token
|
73 |
+
):
|
74 |
+
result["input_ids"].append(self.tokenizer.eos_token_id)
|
75 |
+
result["attention_mask"].append(1)
|
76 |
+
|
77 |
+
if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
|
78 |
+
result["input_ids"] = result["input_ids"][1:]
|
79 |
+
result["attention_mask"] = result["attention_mask"][1:]
|
80 |
+
|
81 |
+
result["labels"] = result["input_ids"].copy()
|
82 |
+
return result
|
83 |
+
|
84 |
|
85 |
class InstructionPromptTokenizingStrategy(PromptTokenizingStrategy):
|
86 |
"""
|
|
|
129 |
)
|
130 |
)
|
131 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
|
133 |
class AlpacaPromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
|
134 |
"""
|
|
|
295 |
)
|
296 |
)
|
297 |
|
298 |
+
def _tokenize(self, prompt, add_eos_token=True, strip_bos_token=False):
|
299 |
result = self.tokenizer(
|
300 |
prompt,
|
301 |
truncation=True,
|
|
|
339 |
return prompt["conversations"]
|
340 |
|
341 |
def tokenize_prompt(self, prompt):
|
342 |
+
result, current_len = tokenize_prompt_default()
|
|
|
|
|
|
|
|
|
|
|
343 |
user_token = self._get_user_token()
|
344 |
assistant_token = self._get_assistant_token()
|
345 |
try:
|
|
|
377 |
)
|
378 |
# everything from this is masked out from the labels
|
379 |
labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
|
380 |
+
|
381 |
+
# pylint: disable=duplicate-code
|
382 |
+
result, current_len = parse_tokenized_to_result(
|
383 |
+
result,
|
384 |
+
current_len,
|
385 |
+
res,
|
386 |
+
labels,
|
387 |
+
pad_token_id=self.tokenizer.pad_token_id,
|
388 |
+
)
|
389 |
return result
|
390 |
except (KeyError, AssertionError, IndexError) as err:
|
391 |
raise InvalidDataException(str(err)) from err
|
|
|
412 |
|
413 |
result["labels"] = result["input_ids"].copy()
|
414 |
return result
|
415 |
+
|
416 |
+
|
417 |
+
def tokenize_prompt_default() -> Tuple[Dict[str, List[int]], int]:
|
418 |
+
"""
|
419 |
+
Returns the default values for the tokenize prompt function
|
420 |
+
"""
|
421 |
+
|
422 |
+
result = {
|
423 |
+
"input_ids": [],
|
424 |
+
"attention_mask": [],
|
425 |
+
"labels": [],
|
426 |
+
}
|
427 |
+
current_len = 0
|
428 |
+
return result, current_len
|
429 |
+
|
430 |
+
|
431 |
+
def parse_tokenized_to_result(
|
432 |
+
result: Dict[str, List[int]],
|
433 |
+
current_len: int,
|
434 |
+
res: Dict[str, List[int]],
|
435 |
+
labels: list[int],
|
436 |
+
pad_token_id: int | None = None,
|
437 |
+
) -> Tuple[Dict[str, List[int]], int]:
|
438 |
+
"""
|
439 |
+
Parses the tokenized prompt and append the tokenized input_ids, attention_mask and labels to the result
|
440 |
+
"""
|
441 |
+
|
442 |
+
input_ids = res["input_ids"]
|
443 |
+
input_len = len(input_ids)
|
444 |
+
result["input_ids"][current_len : current_len + input_len] = input_ids
|
445 |
+
result["attention_mask"][current_len : current_len + input_len] = [
|
446 |
+
1 if x != pad_token_id else 0 for x in input_ids
|
447 |
+
]
|
448 |
+
result["labels"][current_len : current_len + input_len] = labels
|
449 |
+
current_len += input_len
|
450 |
+
|
451 |
+
return result, current_len
|