update sharegpt conversations when chatml chat template is set (#1075) [skip ci]
Browse files* update sharegpt conversations when chatml chat template is set
* add info log when updating sharegpt/chatml conversation
- src/axolotl/cli/__init__.py +7 -1
- src/axolotl/utils/config.py +15 -0
- tests/test_normalize_config.py +24 -1
src/axolotl/cli/__init__.py
CHANGED
@@ -25,7 +25,11 @@ from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
|
|
25 |
from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
|
26 |
from axolotl.logging_config import configure_logging
|
27 |
from axolotl.train import TrainDatasetMeta
|
28 |
-
from axolotl.utils.config import
|
|
|
|
|
|
|
|
|
29 |
from axolotl.utils.data import prepare_dataset
|
30 |
from axolotl.utils.dict import DictDefault
|
31 |
from axolotl.utils.distributed import is_main_process
|
@@ -289,6 +293,8 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
|
|
289 |
|
290 |
normalize_config(cfg)
|
291 |
|
|
|
|
|
292 |
setup_wandb_env_vars(cfg)
|
293 |
|
294 |
setup_mlflow_env_vars(cfg)
|
|
|
25 |
from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
|
26 |
from axolotl.logging_config import configure_logging
|
27 |
from axolotl.train import TrainDatasetMeta
|
28 |
+
from axolotl.utils.config import (
|
29 |
+
normalize_cfg_datasets,
|
30 |
+
normalize_config,
|
31 |
+
validate_config,
|
32 |
+
)
|
33 |
from axolotl.utils.data import prepare_dataset
|
34 |
from axolotl.utils.dict import DictDefault
|
35 |
from axolotl.utils.distributed import is_main_process
|
|
|
293 |
|
294 |
normalize_config(cfg)
|
295 |
|
296 |
+
normalize_cfg_datasets(cfg)
|
297 |
+
|
298 |
setup_wandb_env_vars(cfg)
|
299 |
|
300 |
setup_mlflow_env_vars(cfg)
|
src/axolotl/utils/config.py
CHANGED
@@ -150,6 +150,21 @@ def normalize_config(cfg):
|
|
150 |
log_gpu_memory_usage(LOG, "baseline", cfg.device)
|
151 |
|
152 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
153 |
def validate_config(cfg):
|
154 |
"""
|
155 |
This is a "pre-validation" step that handles the yaml configuration before we have any
|
|
|
150 |
log_gpu_memory_usage(LOG, "baseline", cfg.device)
|
151 |
|
152 |
|
153 |
+
def normalize_cfg_datasets(cfg):
|
154 |
+
"""
|
155 |
+
helpers for mapping chat_template to various dataset configurations as necessary
|
156 |
+
"""
|
157 |
+
|
158 |
+
if cfg.chat_template and cfg.chat_template == "chatml":
|
159 |
+
if cfg.datasets:
|
160 |
+
for idx, ds_cfg in enumerate(cfg.datasets):
|
161 |
+
if ds_cfg.type == "sharegpt" and not ds_cfg.conversation:
|
162 |
+
LOG.info(
|
163 |
+
f"updating dataset {ds_cfg.path} with `conversation: chatml` to match your chat_template"
|
164 |
+
)
|
165 |
+
cfg.datasets[idx].conversation = "chatml"
|
166 |
+
|
167 |
+
|
168 |
def validate_config(cfg):
|
169 |
"""
|
170 |
This is a "pre-validation" step that handles the yaml configuration before we have any
|
tests/test_normalize_config.py
CHANGED
@@ -3,7 +3,7 @@ Test classes for checking functionality of the cfg normalization
|
|
3 |
"""
|
4 |
import unittest
|
5 |
|
6 |
-
from axolotl.utils.config import normalize_config
|
7 |
from axolotl.utils.dict import DictDefault
|
8 |
|
9 |
|
@@ -44,3 +44,26 @@ class NormalizeConfigTestCase(unittest.TestCase):
|
|
44 |
normalize_config(cfg)
|
45 |
|
46 |
assert cfg.base_model_config == cfg.base_model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
"""
|
4 |
import unittest
|
5 |
|
6 |
+
from axolotl.utils.config import normalize_cfg_datasets, normalize_config
|
7 |
from axolotl.utils.dict import DictDefault
|
8 |
|
9 |
|
|
|
44 |
normalize_config(cfg)
|
45 |
|
46 |
assert cfg.base_model_config == cfg.base_model
|
47 |
+
|
48 |
+
def test_chat_template_chatml(self):
|
49 |
+
cfg = DictDefault(
|
50 |
+
{
|
51 |
+
"chat_template": "chatml",
|
52 |
+
"datasets": [
|
53 |
+
{
|
54 |
+
"path": "lorem/ipsum",
|
55 |
+
"type": "sharegpt",
|
56 |
+
"conversation": "vicuna_v1.1",
|
57 |
+
},
|
58 |
+
{
|
59 |
+
"path": "sit/amet",
|
60 |
+
"type": "sharegpt",
|
61 |
+
},
|
62 |
+
],
|
63 |
+
}
|
64 |
+
)
|
65 |
+
|
66 |
+
normalize_cfg_datasets(cfg)
|
67 |
+
|
68 |
+
assert cfg.datasets[0].conversation == "vicuna_v1.1"
|
69 |
+
assert cfg.datasets[1].conversation == "chatml"
|