Support user-defined prompt processing strategies for dpo (#1248)
Browse files* support user-defined prompt processing strategies for dpo
* interpret dict dataset types as user-defined
* fix lint errors
* setup pydantic config for validation of User defined DPO
---------
Co-authored-by: Wing Lian <wing.lian@gmail.com>
- src/axolotl/prompt_strategies/dpo/__init__.py +2 -3
- src/axolotl/prompt_strategies/dpo/chatml.py +7 -3
- src/axolotl/prompt_strategies/dpo/user_defined.py +41 -0
- src/axolotl/prompt_strategies/dpo/zephyr.py +1 -1
- src/axolotl/utils/config/models/input/v0_4_1/__init__.py +13 -1
- src/axolotl/utils/data.py +3 -1
src/axolotl/prompt_strategies/dpo/__init__.py
CHANGED
@@ -8,14 +8,13 @@ import logging
|
|
8 |
LOG = logging.getLogger("axolotl")
|
9 |
|
10 |
|
11 |
-
def load(strategy, cfg):
|
12 |
try:
|
13 |
load_fn = strategy.split(".")[-1]
|
14 |
strategy = ".".join(strategy.split(".")[:-1])
|
15 |
mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies.dpo")
|
16 |
func = getattr(mod, load_fn)
|
17 |
-
|
18 |
-
return func(cfg, **load_kwargs)
|
19 |
except Exception: # pylint: disable=broad-exception-caught
|
20 |
LOG.warning(f"unable to load strategy {strategy}")
|
21 |
return None
|
|
|
8 |
LOG = logging.getLogger("axolotl")
|
9 |
|
10 |
|
11 |
+
def load(strategy, cfg, **kwargs):
|
12 |
try:
|
13 |
load_fn = strategy.split(".")[-1]
|
14 |
strategy = ".".join(strategy.split(".")[:-1])
|
15 |
mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies.dpo")
|
16 |
func = getattr(mod, load_fn)
|
17 |
+
return func(cfg, **kwargs)
|
|
|
18 |
except Exception: # pylint: disable=broad-exception-caught
|
19 |
LOG.warning(f"unable to load strategy {strategy}")
|
20 |
return None
|
src/axolotl/prompt_strategies/dpo/chatml.py
CHANGED
@@ -5,6 +5,7 @@ DPO strategies for chatml
|
|
5 |
|
6 |
def argilla(
|
7 |
cfg,
|
|
|
8 |
): # pylint: disable=possibly-unused-variable,unused-argument
|
9 |
def transform_fn(sample):
|
10 |
if "system" in sample and sample["system"]:
|
@@ -25,6 +26,7 @@ def argilla(
|
|
25 |
|
26 |
def icr(
|
27 |
cfg,
|
|
|
28 |
): # pylint: disable=possibly-unused-variable,unused-argument
|
29 |
"""
|
30 |
chatml transforms for datasets with system, input, chosen, rejected
|
@@ -48,7 +50,7 @@ def icr(
|
|
48 |
return transform_fn
|
49 |
|
50 |
|
51 |
-
def intel(cfg): # pylint: disable=possibly-unused-variable,unused-argument
|
52 |
"""
|
53 |
For Intel Orca DPO Pairs
|
54 |
"""
|
@@ -70,7 +72,9 @@ def intel(cfg): # pylint: disable=possibly-unused-variable,unused-argument
|
|
70 |
return transform_fn
|
71 |
|
72 |
|
73 |
-
def prompt_pairs(
|
|
|
|
|
74 |
def transform_fn(sample):
|
75 |
if "system" in sample and sample["system"]:
|
76 |
sample["prompt"] = (
|
@@ -88,7 +92,7 @@ def prompt_pairs(cfg): # pylint: disable=possibly-unused-variable,unused-argume
|
|
88 |
return transform_fn
|
89 |
|
90 |
|
91 |
-
def ultra(cfg): # pylint: disable=possibly-unused-variable,unused-argument
|
92 |
"""
|
93 |
for ultrafeedback binarized conversations
|
94 |
"""
|
|
|
5 |
|
6 |
def argilla(
|
7 |
cfg,
|
8 |
+
**kwargs,
|
9 |
): # pylint: disable=possibly-unused-variable,unused-argument
|
10 |
def transform_fn(sample):
|
11 |
if "system" in sample and sample["system"]:
|
|
|
26 |
|
27 |
def icr(
|
28 |
cfg,
|
29 |
+
**kwargs,
|
30 |
): # pylint: disable=possibly-unused-variable,unused-argument
|
31 |
"""
|
32 |
chatml transforms for datasets with system, input, chosen, rejected
|
|
|
50 |
return transform_fn
|
51 |
|
52 |
|
53 |
+
def intel(cfg, **kwargs): # pylint: disable=possibly-unused-variable,unused-argument
|
54 |
"""
|
55 |
For Intel Orca DPO Pairs
|
56 |
"""
|
|
|
72 |
return transform_fn
|
73 |
|
74 |
|
75 |
+
def prompt_pairs(
|
76 |
+
cfg, **kwargs
|
77 |
+
): # pylint: disable=possibly-unused-variable,unused-argument
|
78 |
def transform_fn(sample):
|
79 |
if "system" in sample and sample["system"]:
|
80 |
sample["prompt"] = (
|
|
|
92 |
return transform_fn
|
93 |
|
94 |
|
95 |
+
def ultra(cfg, **kwargs): # pylint: disable=possibly-unused-variable,unused-argument
|
96 |
"""
|
97 |
for ultrafeedback binarized conversations
|
98 |
"""
|
src/axolotl/prompt_strategies/dpo/user_defined.py
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
User-defined DPO strategies
|
3 |
+
"""
|
4 |
+
|
5 |
+
|
6 |
+
def default(cfg, dataset_idx=0, **kwargs): # pylint: disable=unused-argument
|
7 |
+
ds_cfg = cfg["datasets"][dataset_idx]["type"]
|
8 |
+
if not isinstance(ds_cfg, dict):
|
9 |
+
raise ValueError(
|
10 |
+
f"User-defined dataset type must be a dictionary. Got: {ds_cfg}"
|
11 |
+
)
|
12 |
+
field_prompt = ds_cfg.get("field_prompt", "prompt")
|
13 |
+
field_system = ds_cfg.get("field_system", "system")
|
14 |
+
field_chosen = ds_cfg.get("field_chosen", "chosen")
|
15 |
+
field_rejected = ds_cfg.get("field_rejected", "rejected")
|
16 |
+
prompt_format = ds_cfg.get("prompt_format")
|
17 |
+
if not prompt_format:
|
18 |
+
prompt_format = "{" + field_prompt + "}"
|
19 |
+
chosen_format = ds_cfg.get("chosen_format")
|
20 |
+
if not chosen_format:
|
21 |
+
chosen_format = "{" + field_chosen + "}"
|
22 |
+
rejected_format = ds_cfg.get("rejected_format")
|
23 |
+
if not rejected_format:
|
24 |
+
rejected_format = "{" + field_rejected + "}"
|
25 |
+
|
26 |
+
def transform_fn(sample):
|
27 |
+
if (
|
28 |
+
"{" + field_system + "}" in prompt_format
|
29 |
+
and field_system in sample
|
30 |
+
and sample[field_system]
|
31 |
+
):
|
32 |
+
sample["prompt"] = prompt_format.format(
|
33 |
+
system=sample[field_system], prompt=sample[field_prompt]
|
34 |
+
)
|
35 |
+
else:
|
36 |
+
sample["prompt"] = prompt_format.format(prompt=sample["prompt"])
|
37 |
+
sample["chosen"] = chosen_format.format(chosen=sample[field_chosen])
|
38 |
+
sample["rejected"] = rejected_format.format(rejected=sample[field_rejected])
|
39 |
+
return sample
|
40 |
+
|
41 |
+
return transform_fn
|
src/axolotl/prompt_strategies/dpo/zephyr.py
CHANGED
@@ -3,7 +3,7 @@ DPO strategies for zephyr
|
|
3 |
"""
|
4 |
|
5 |
|
6 |
-
def nectar(cfg): # pylint: disable=possibly-unused-variable,unused-argument
|
7 |
def transform_fn(sample):
|
8 |
data = {}
|
9 |
data["prompt"] = (
|
|
|
3 |
"""
|
4 |
|
5 |
|
6 |
+
def nectar(cfg, **kwargs): # pylint: disable=possibly-unused-variable,unused-argument
|
7 |
def transform_fn(sample):
|
8 |
data = {}
|
9 |
data["prompt"] = (
|
src/axolotl/utils/config/models/input/v0_4_1/__init__.py
CHANGED
@@ -85,12 +85,24 @@ class SFTDataset(BaseModel):
|
|
85 |
field_model: Optional[str] = None
|
86 |
|
87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
class DPODataset(BaseModel):
|
89 |
"""DPO configuration subset"""
|
90 |
|
91 |
path: Optional[str] = None
|
92 |
split: Optional[str] = None
|
93 |
-
type: Optional[str] = None
|
94 |
data_files: Optional[List[str]] = None
|
95 |
|
96 |
|
|
|
85 |
field_model: Optional[str] = None
|
86 |
|
87 |
|
88 |
+
class UserDefinedDPOType(BaseModel):
|
89 |
+
"""User defined typing for DPO"""
|
90 |
+
|
91 |
+
field_system: Optional[str] = None
|
92 |
+
field_prompt: Optional[str] = None
|
93 |
+
field_chosen: Optional[str] = None
|
94 |
+
field_rejected: Optional[str] = None
|
95 |
+
prompt_format: Optional[str] = None
|
96 |
+
chosen_format: Optional[str] = None
|
97 |
+
rejected_format: Optional[str] = None
|
98 |
+
|
99 |
+
|
100 |
class DPODataset(BaseModel):
|
101 |
"""DPO configuration subset"""
|
102 |
|
103 |
path: Optional[str] = None
|
104 |
split: Optional[str] = None
|
105 |
+
type: Optional[Union[UserDefinedDPOType, str]] = None
|
106 |
data_files: Optional[List[str]] = None
|
107 |
|
108 |
|
src/axolotl/utils/data.py
CHANGED
@@ -937,7 +937,9 @@ def load_prepare_dpo_datasets(cfg):
|
|
937 |
for i, data_set in enumerate(split_datasets):
|
938 |
_type = dataset_cfgs[i]["type"]
|
939 |
if _type:
|
940 |
-
|
|
|
|
|
941 |
split_datasets[i] = data_set.map(
|
942 |
ds_transform_fn,
|
943 |
desc="Mapping RL Dataset",
|
|
|
937 |
for i, data_set in enumerate(split_datasets):
|
938 |
_type = dataset_cfgs[i]["type"]
|
939 |
if _type:
|
940 |
+
if isinstance(_type, DictDefault):
|
941 |
+
_type = "user_defined.default"
|
942 |
+
ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
|
943 |
split_datasets[i] = data_set.map(
|
944 |
ds_transform_fn,
|
945 |
desc="Mapping RL Dataset",
|