support for explicit test_dataset definition for evals (#786)
Browse files- src/axolotl/utils/config.py +5 -0
- src/axolotl/utils/data.py +39 -29
src/axolotl/utils/config.py
CHANGED
@@ -519,6 +519,11 @@ def validate_config(cfg):
|
|
519 |
"bf16.enabled or fp16.enabled must be set to true when using ZeRO-3 with flash-attention"
|
520 |
)
|
521 |
|
|
|
|
|
|
|
|
|
|
|
522 |
# TODO
|
523 |
# MPT 7b
|
524 |
# https://github.com/facebookresearch/bitsandbytes/issues/25
|
|
|
519 |
"bf16.enabled or fp16.enabled must be set to true when using ZeRO-3 with flash-attention"
|
520 |
)
|
521 |
|
522 |
+
if cfg.test_datasets and cfg.val_set_size:
|
523 |
+
raise ValueError(
|
524 |
+
"non-zero val_set_size should not be used with test_datasets configuration"
|
525 |
+
)
|
526 |
+
|
527 |
# TODO
|
528 |
# MPT 7b
|
529 |
# https://github.com/facebookresearch/bitsandbytes/issues/25
|
src/axolotl/utils/data.py
CHANGED
@@ -4,7 +4,7 @@ import hashlib
|
|
4 |
import logging
|
5 |
from collections import defaultdict
|
6 |
from pathlib import Path
|
7 |
-
from typing import Dict, List, Tuple, Union
|
8 |
|
9 |
import torch
|
10 |
from datasets import (
|
@@ -65,9 +65,17 @@ def prepare_dataset(cfg, tokenizer):
|
|
65 |
prompters = []
|
66 |
if not cfg.pretraining_dataset:
|
67 |
with zero_first(is_main_process()):
|
68 |
-
|
69 |
-
|
70 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
else:
|
72 |
path = cfg.pretraining_dataset
|
73 |
name = None
|
@@ -108,8 +116,12 @@ def prepare_dataset(cfg, tokenizer):
|
|
108 |
|
109 |
|
110 |
def load_tokenized_prepared_datasets(
|
111 |
-
tokenizer,
|
|
|
|
|
|
|
112 |
) -> Tuple[DatasetDict, List[Prompter]]:
|
|
|
113 |
tokenizer_name = tokenizer.__class__.__name__
|
114 |
ds_hash = str(
|
115 |
md5(
|
@@ -126,7 +138,7 @@ def load_tokenized_prepared_datasets(
|
|
126 |
sorted(
|
127 |
[
|
128 |
f"{d.path}:{d.type}:{d.shards}:{d.conversation}"
|
129 |
-
for d in
|
130 |
]
|
131 |
)
|
132 |
)
|
@@ -149,7 +161,7 @@ def load_tokenized_prepared_datasets(
|
|
149 |
f"{cfg.push_dataset_to_hub}/{ds_hash}",
|
150 |
token=use_auth_token,
|
151 |
)
|
152 |
-
dataset = dataset[
|
153 |
except Exception: # pylint: disable=broad-except # nosec
|
154 |
pass
|
155 |
|
@@ -188,8 +200,8 @@ def load_tokenized_prepared_datasets(
|
|
188 |
yield dataset
|
189 |
|
190 |
# pylint: disable=invalid-name
|
191 |
-
for config_dataset in for_d_in_datasets(
|
192 |
-
ds: Union[Dataset, DatasetDict] = None
|
193 |
ds_from_hub = False
|
194 |
try:
|
195 |
load_dataset(
|
@@ -342,16 +354,6 @@ def load_tokenized_prepared_datasets(
|
|
342 |
)
|
343 |
if not ds:
|
344 |
raise ValueError("unhandled dataset load")
|
345 |
-
# support for using a subset of the data
|
346 |
-
if config_dataset.shards:
|
347 |
-
if "train" in ds:
|
348 |
-
ds = ds.shuffle(seed=seed)["train"].shard(
|
349 |
-
num_shards=config_dataset.shards, index=0
|
350 |
-
)
|
351 |
-
else:
|
352 |
-
ds = ds.shuffle(seed=seed).shard(
|
353 |
-
num_shards=config_dataset.shards, index=0
|
354 |
-
)
|
355 |
|
356 |
d_base_type = d_prompt_style = None
|
357 |
d_type = config_dataset.type
|
@@ -359,17 +361,21 @@ def load_tokenized_prepared_datasets(
|
|
359 |
d_type_split = d_type.split(":")
|
360 |
d_base_type = d_type_split[0]
|
361 |
d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
and config_dataset.train_on_split in ds
|
368 |
-
):
|
369 |
-
ds = ds[config_dataset.train_on_split]
|
370 |
elif isinstance(ds, DatasetDict):
|
371 |
raise ValueError(
|
372 |
-
f"no
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
373 |
)
|
374 |
|
375 |
dataset_wrapper, dataset_prompter = get_dataset_wrapper(
|
@@ -428,6 +434,7 @@ def load_prepare_datasets(
|
|
428 |
tokenizer: PreTrainedTokenizerBase,
|
429 |
cfg,
|
430 |
default_dataset_prepared_path,
|
|
|
431 |
) -> Tuple[Dataset, Dataset, List[Prompter]]:
|
432 |
dataset, prompters = load_tokenized_prepared_datasets(
|
433 |
tokenizer, cfg, default_dataset_prepared_path
|
@@ -442,7 +449,7 @@ def load_prepare_datasets(
|
|
442 |
index=cfg.dataset_shard_idx,
|
443 |
)
|
444 |
|
445 |
-
if cfg.val_set_size:
|
446 |
# ensure we end up with the same fingerprint by doing rank0 first and being able to cache
|
447 |
to_hash_train = (
|
448 |
dataset._fingerprint # pylint: disable=protected-access
|
@@ -475,6 +482,9 @@ def load_prepare_datasets(
|
|
475 |
|
476 |
train_dataset = dataset["train"]
|
477 |
eval_dataset = dataset["test"]
|
|
|
|
|
|
|
478 |
else:
|
479 |
train_dataset = dataset
|
480 |
eval_dataset = None
|
|
|
4 |
import logging
|
5 |
from collections import defaultdict
|
6 |
from pathlib import Path
|
7 |
+
from typing import Dict, List, Optional, Tuple, Union
|
8 |
|
9 |
import torch
|
10 |
from datasets import (
|
|
|
65 |
prompters = []
|
66 |
if not cfg.pretraining_dataset:
|
67 |
with zero_first(is_main_process()):
|
68 |
+
if cfg.test_datasets:
|
69 |
+
train_dataset, _, prompters = load_prepare_datasets(
|
70 |
+
tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH, split="train"
|
71 |
+
)
|
72 |
+
_, eval_dataset, _ = load_prepare_datasets(
|
73 |
+
tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH, split="test"
|
74 |
+
)
|
75 |
+
else:
|
76 |
+
train_dataset, eval_dataset, prompters = load_prepare_datasets(
|
77 |
+
tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
|
78 |
+
)
|
79 |
else:
|
80 |
path = cfg.pretraining_dataset
|
81 |
name = None
|
|
|
116 |
|
117 |
|
118 |
def load_tokenized_prepared_datasets(
|
119 |
+
tokenizer,
|
120 |
+
cfg,
|
121 |
+
default_dataset_prepared_path,
|
122 |
+
split="train",
|
123 |
) -> Tuple[DatasetDict, List[Prompter]]:
|
124 |
+
cfg_datasets = cfg.test_datasets if split == "test" else cfg.datasets
|
125 |
tokenizer_name = tokenizer.__class__.__name__
|
126 |
ds_hash = str(
|
127 |
md5(
|
|
|
138 |
sorted(
|
139 |
[
|
140 |
f"{d.path}:{d.type}:{d.shards}:{d.conversation}"
|
141 |
+
for d in cfg_datasets
|
142 |
]
|
143 |
)
|
144 |
)
|
|
|
161 |
f"{cfg.push_dataset_to_hub}/{ds_hash}",
|
162 |
token=use_auth_token,
|
163 |
)
|
164 |
+
dataset = dataset[split]
|
165 |
except Exception: # pylint: disable=broad-except # nosec
|
166 |
pass
|
167 |
|
|
|
200 |
yield dataset
|
201 |
|
202 |
# pylint: disable=invalid-name
|
203 |
+
for config_dataset in for_d_in_datasets(cfg_datasets):
|
204 |
+
ds: Optional[Union[Dataset, DatasetDict]] = None
|
205 |
ds_from_hub = False
|
206 |
try:
|
207 |
load_dataset(
|
|
|
354 |
)
|
355 |
if not ds:
|
356 |
raise ValueError("unhandled dataset load")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
357 |
|
358 |
d_base_type = d_prompt_style = None
|
359 |
d_type = config_dataset.type
|
|
|
361 |
d_type_split = d_type.split(":")
|
362 |
d_base_type = d_type_split[0]
|
363 |
d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
|
364 |
+
|
365 |
+
if config_dataset.split and config_dataset.split in ds:
|
366 |
+
ds = ds[config_dataset.split]
|
367 |
+
elif split in ds:
|
368 |
+
ds = ds[split]
|
|
|
|
|
|
|
369 |
elif isinstance(ds, DatasetDict):
|
370 |
raise ValueError(
|
371 |
+
f"no {split} split found for dataset {config_dataset.path}, you may specify a split with 'split: `"
|
372 |
+
)
|
373 |
+
|
374 |
+
# support for using a subset of the data
|
375 |
+
if config_dataset.shards:
|
376 |
+
shards_idx = config_dataset.get("shards_idx", 0)
|
377 |
+
ds = ds.shuffle(seed=seed).shard(
|
378 |
+
num_shards=config_dataset.shards, index=shards_idx
|
379 |
)
|
380 |
|
381 |
dataset_wrapper, dataset_prompter = get_dataset_wrapper(
|
|
|
434 |
tokenizer: PreTrainedTokenizerBase,
|
435 |
cfg,
|
436 |
default_dataset_prepared_path,
|
437 |
+
split="train",
|
438 |
) -> Tuple[Dataset, Dataset, List[Prompter]]:
|
439 |
dataset, prompters = load_tokenized_prepared_datasets(
|
440 |
tokenizer, cfg, default_dataset_prepared_path
|
|
|
449 |
index=cfg.dataset_shard_idx,
|
450 |
)
|
451 |
|
452 |
+
if split == "train" and cfg.val_set_size:
|
453 |
# ensure we end up with the same fingerprint by doing rank0 first and being able to cache
|
454 |
to_hash_train = (
|
455 |
dataset._fingerprint # pylint: disable=protected-access
|
|
|
482 |
|
483 |
train_dataset = dataset["train"]
|
484 |
eval_dataset = dataset["test"]
|
485 |
+
elif split == "test":
|
486 |
+
train_dataset = None
|
487 |
+
eval_dataset = dataset
|
488 |
else:
|
489 |
train_dataset = dataset
|
490 |
eval_dataset = None
|