plain input/output prompt strategy w/o chat templates (#1346)
Browse files* plain input/output prompt strategy w/o chat templates
* disable duplicate code check
* make sure to add an eos/eot token to the end of the output so it will stop
* multi turn segement support and test
src/axolotl/prompt_strategies/input_output.py
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Module for plain input/output prompt pairs"""
|
| 2 |
+
from typing import Generator, Tuple
|
| 3 |
+
|
| 4 |
+
from axolotl.prompt_tokenizers import PromptTokenizingStrategy
|
| 5 |
+
from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
class RawInputOutputStrategy(PromptTokenizingStrategy):
|
| 9 |
+
"""Prompt Strategy class for input/output pairs"""
|
| 10 |
+
|
| 11 |
+
def __init__(self, *args, eos_token=None, **kwargs):
|
| 12 |
+
super().__init__(*args, **kwargs)
|
| 13 |
+
self.eos_token = eos_token
|
| 14 |
+
if not eos_token:
|
| 15 |
+
self.eos_token = self.tokenizer.eos_token
|
| 16 |
+
|
| 17 |
+
def tokenize_prompt(self, prompt):
|
| 18 |
+
# pylint: disable=duplicate-code
|
| 19 |
+
input_ids = []
|
| 20 |
+
labels = []
|
| 21 |
+
for label, text in self.prompter.build_prompt(prompt["segments"]):
|
| 22 |
+
tokenized_output = self.tokenizer(
|
| 23 |
+
text, add_special_tokens=False, return_tensors=None
|
| 24 |
+
)["input_ids"]
|
| 25 |
+
input_ids += tokenized_output
|
| 26 |
+
if label or self.train_on_inputs:
|
| 27 |
+
labels += tokenized_output
|
| 28 |
+
else:
|
| 29 |
+
labels += [IGNORE_TOKEN_ID] * len(tokenized_output)
|
| 30 |
+
|
| 31 |
+
tokenized_prompt = {
|
| 32 |
+
"input_ids": input_ids,
|
| 33 |
+
"labels": labels,
|
| 34 |
+
"attention_mask": [1] * len(input_ids),
|
| 35 |
+
}
|
| 36 |
+
|
| 37 |
+
return tokenized_prompt
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
class RawInputOutputPrompter(Prompter):
|
| 41 |
+
"""prompter for raw i/o data"""
|
| 42 |
+
|
| 43 |
+
def build_prompt(self, source) -> Generator[Tuple[bool, str], None, None]:
|
| 44 |
+
for segment in source:
|
| 45 |
+
yield segment["label"], segment["text"]
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
def load(tokenizer, cfg):
|
| 49 |
+
return RawInputOutputStrategy(
|
| 50 |
+
RawInputOutputPrompter(),
|
| 51 |
+
tokenizer,
|
| 52 |
+
cfg.train_on_inputs,
|
| 53 |
+
cfg.sequence_len,
|
| 54 |
+
)
|
tests/prompt_strategies/test_raw_io.py
ADDED
|
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Test module for raw i/o data for prompts
|
| 3 |
+
"""
|
| 4 |
+
import pytest
|
| 5 |
+
from datasets import Dataset
|
| 6 |
+
from tokenizers import AddedToken
|
| 7 |
+
from transformers import AutoTokenizer
|
| 8 |
+
|
| 9 |
+
from axolotl.datasets import TokenizedPromptDataset
|
| 10 |
+
from axolotl.prompt_strategies.input_output import (
|
| 11 |
+
RawInputOutputPrompter,
|
| 12 |
+
RawInputOutputStrategy,
|
| 13 |
+
)
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
@pytest.fixture(name="segments_dataset")
|
| 17 |
+
def fixture_sharegpt_dataset():
|
| 18 |
+
return Dataset.from_list(
|
| 19 |
+
[
|
| 20 |
+
{
|
| 21 |
+
"segments": [
|
| 22 |
+
{
|
| 23 |
+
"label": False,
|
| 24 |
+
"text": "<s>hello ",
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"label": True,
|
| 28 |
+
"text": "hi there.<eot>",
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"label": False,
|
| 32 |
+
"text": "goodbye ",
|
| 33 |
+
},
|
| 34 |
+
{
|
| 35 |
+
"label": True,
|
| 36 |
+
"text": "farewell<eot>",
|
| 37 |
+
},
|
| 38 |
+
]
|
| 39 |
+
}
|
| 40 |
+
]
|
| 41 |
+
)
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
@pytest.fixture(name="tokenizer")
|
| 45 |
+
def fixture_tokenizer():
|
| 46 |
+
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
|
| 47 |
+
tokenizer.add_tokens(
|
| 48 |
+
[
|
| 49 |
+
AddedToken("<eot>", rstrip=False, lstrip=False, normalized=False),
|
| 50 |
+
]
|
| 51 |
+
)
|
| 52 |
+
|
| 53 |
+
return tokenizer
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
class TestRawInputOutputPrompts:
|
| 57 |
+
"""
|
| 58 |
+
Test class for raw i/o prompter
|
| 59 |
+
"""
|
| 60 |
+
|
| 61 |
+
def test_segment_prompts(self, segments_dataset, tokenizer):
|
| 62 |
+
strategy = RawInputOutputStrategy(
|
| 63 |
+
RawInputOutputPrompter(),
|
| 64 |
+
tokenizer,
|
| 65 |
+
False, # train_on_inputs
|
| 66 |
+
2048, # sequence_len
|
| 67 |
+
)
|
| 68 |
+
|
| 69 |
+
dataset_wrapper = TokenizedPromptDataset(
|
| 70 |
+
strategy, segments_dataset, process_count=1
|
| 71 |
+
)
|
| 72 |
+
|
| 73 |
+
input_ids = dataset_wrapper[0]["input_ids"]
|
| 74 |
+
labels = dataset_wrapper[0]["labels"]
|
| 75 |
+
|
| 76 |
+
assert (
|
| 77 |
+
tokenizer.decode(input_ids)
|
| 78 |
+
== "<s> hello hi there.<eot> goodbye farewell<eot>"
|
| 79 |
+
)
|
| 80 |
+
# fmt: off
|
| 81 |
+
assert input_ids == [
|
| 82 |
+
1, # <s>
|
| 83 |
+
6312, # hell
|
| 84 |
+
28709, # o
|
| 85 |
+
28705, #
|
| 86 |
+
12014, # hi
|
| 87 |
+
736, # there
|
| 88 |
+
28723, # .
|
| 89 |
+
32000, # <eot>
|
| 90 |
+
1179, # good
|
| 91 |
+
17664, # bye
|
| 92 |
+
28705, #
|
| 93 |
+
19111, # fare
|
| 94 |
+
5458, # well
|
| 95 |
+
32000, # <eot>
|
| 96 |
+
]
|
| 97 |
+
# fmt: on
|
| 98 |
+
|
| 99 |
+
# fmt: off
|
| 100 |
+
assert labels == [
|
| 101 |
+
-100, # <s>
|
| 102 |
+
-100, # hell
|
| 103 |
+
-100, # o
|
| 104 |
+
-100, #
|
| 105 |
+
12014, # hi
|
| 106 |
+
736, # there
|
| 107 |
+
28723, # .
|
| 108 |
+
32000, # <eot>
|
| 109 |
+
-100, # good
|
| 110 |
+
-100, # bye
|
| 111 |
+
-100, #
|
| 112 |
+
19111, # fare
|
| 113 |
+
5458, # well
|
| 114 |
+
32000, # <eot>
|
| 115 |
+
]
|
| 116 |
+
# fmt: on
|