forked from openai/evals
-
Notifications
You must be signed in to change notification settings - Fork 0
/
registry.py
274 lines (226 loc) · 9.38 KB
/
registry.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
"""
Functions to handle registration of evals. To add a new eval to the registry,
add an entry in one of the YAML files in the `../registry` dir.
By convention, every eval name should start with {base_eval}.{split}.
"""
import copy
import difflib
import functools
import logging
import os
import re
from functools import cached_property
from pathlib import Path
from typing import Any, Iterator, Optional, Sequence, Type, Union
import openai
import yaml
from evals import OpenAIChatCompletionFn, OpenAICompletionFn
from evals.api import CompletionFn, DummyCompletionFn
from evals.base import BaseEvalSpec, CompletionFnSpec, EvalSetSpec, EvalSpec
from evals.elsuite.modelgraded.base import ModelGradedSpec
from evals.utils.misc import make_object
logger = logging.getLogger(__name__)
DEFAULT_PATHS = [Path(__file__).parents[0].resolve() / "registry", Path.home() / ".evals"]
def n_ctx_from_model_name(model_name: str) -> Optional[int]:
"""Returns n_ctx for a given API model name. Model list last updated 2023-03-14."""
# note that for most models, the max tokens is n_ctx + 1
DICT_OF_N_CTX_BY_MODEL_NAME_PREFIX: dict[str, int] = {
"gpt-3.5-turbo-": 4096,
"gpt-4-": 8192,
"gpt-4-32k-": 32768,
}
DICT_OF_N_CTX_BY_MODEL_NAME: dict[str, int] = {
"ada": 2048,
"text-ada-001": 2048,
"babbage": 2048,
"text-babbage-001": 2048,
"curie": 2048,
"text-curie-001": 2048,
"davinci": 2048,
"text-davinci-001": 2048,
"code-davinci-002": 8000,
"text-davinci-002": 4096,
"text-davinci-003": 4096,
"gpt-3.5-turbo": 4096,
"gpt-3.5-turbo-0301": 4096,
"gpt-4": 8192,
"gpt-4-0314": 8192,
"gpt-4-32k": 32768,
"gpt-4-32k-0314": 32768,
}
# first, look for a prefix match
for model_prefix, n_ctx in DICT_OF_N_CTX_BY_MODEL_NAME_PREFIX.items():
if model_name.startswith(model_prefix):
return n_ctx
# otherwise, look for an exact match and return None if not found
return DICT_OF_N_CTX_BY_MODEL_NAME.get(model_name, None)
class Registry:
def __init__(self, registry_paths: Sequence[Union[str, Path]] = DEFAULT_PATHS):
self._registry_paths = [Path(p) if isinstance(p, str) else p for p in registry_paths]
def add_registry_paths(self, paths: list[Union[str, Path]]):
self._registry_paths.extend([Path(p) if isinstance(p, str) else p for p in paths])
@cached_property
def api_model_ids(self):
return [m["id"] for m in openai.Model.list()["data"]]
def make_completion_fn(self, name: str) -> CompletionFn:
"""
Create a CompletionFn. The name can be one of the following formats:
1. openai-model-id (e.g. "gpt-3.5-turbo")
2. completion-fn-id (from the registry)
"""
if name == "dummy":
return DummyCompletionFn()
n_ctx = n_ctx_from_model_name(name)
CHAT_MODELS = {
"gpt-3.5-turbo",
"gpt-3.5-turbo-0301",
"gpt-4",
"gpt-4-0314",
"gpt-4-32k",
"gpt-4-32k-0314",
}
if name in CHAT_MODELS:
return OpenAIChatCompletionFn(model=name, n_ctx=n_ctx)
elif name in self.api_model_ids:
return OpenAICompletionFn(model=name, n_ctx=n_ctx)
# No match, so try to find a completion-fn-id in the registry
spec = self.get_completion_fn(name)
if spec is None:
raise ValueError(f"Could not find CompletionFn in the registry with ID {name}")
if spec.args is None:
spec.args = {}
spec.args["registry"] = self
instance = make_object(spec.cls)(**spec.args or {})
assert isinstance(instance, CompletionFn), f"{name} must be a CompletionFn"
return instance
def get_class(self, spec: dict) -> Any:
return make_object(spec.cls, **(spec.args if spec.args else {}))
def _dereference(self, name: str, d: dict, object: str, type: Type, **kwargs: dict) -> dict:
if not name in d:
logger.warning(
(
f"{object} '{name}' not found. "
f"Closest matches: {difflib.get_close_matches(name, d.keys(), n=5)}"
)
)
return None
def get_alias():
if isinstance(d[name], str):
return d[name]
if isinstance(d[name], dict) and "id" in d[name]:
return d[name]["id"]
return None
logger.debug(f"Looking for {name}")
while True:
alias = get_alias()
if alias is None:
break
name = alias
spec = d[name]
if kwargs:
spec = copy.deepcopy(spec)
spec.update(kwargs)
try:
return type(**spec)
except TypeError as e:
raise TypeError(f"Error while processing {object} '{name}': {e}")
def get_modelgraded_spec(self, name: str, **kwargs: dict) -> dict[str, Any]:
assert name in self._modelgraded_specs, (
f"Modelgraded spec {name} not found. "
f"Closest matches: {difflib.get_close_matches(name, self._modelgraded_specs.keys(), n=5)}"
)
return self._dereference(
name, self._modelgraded_specs, "modelgraded spec", ModelGradedSpec, **kwargs
)
def get_completion_fn(self, name: str) -> CompletionFnSpec:
return self._dereference(name, self._completion_fns, "completion_fn", CompletionFnSpec)
def get_eval(self, name: str) -> EvalSpec:
return self._dereference(name, self._evals, "eval", EvalSpec)
def get_eval_set(self, name: str) -> EvalSetSpec:
return self._dereference(name, self._eval_sets, "eval set", EvalSetSpec)
def get_evals(self, patterns: Sequence[str]) -> Iterator[EvalSpec]:
# valid patterns: hello, hello.dev*, hello.dev.*-v1
def get_regexp(pattern):
pattern = pattern.replace(".", "\\.")
pattern = pattern.replace("*", ".*")
return re.compile(f"^{pattern}$")
regexps = list(map(get_regexp, patterns))
for name in self._evals:
# if any regexps match, return the name
if any(map(lambda regexp: regexp.match(name), regexps)):
yield self.get_eval(name)
def get_base_evals(self) -> list[BaseEvalSpec]:
base_evals = []
for name, spec in self._evals.items():
if name.count(".") == 0:
base_evals.append(self.get_base_eval(name))
return base_evals
def get_base_eval(self, name: str) -> BaseEvalSpec:
if not name in self._evals:
return None
spec_or_alias = self._evals[name]
if isinstance(spec_or_alias, dict):
spec = spec_or_alias
try:
return BaseEvalSpec(**spec)
except TypeError as e:
raise TypeError(f"Error while processing base eval {name}: {e}")
alias = spec_or_alias
return BaseEvalSpec(id=alias)
def _process_file(self, registry, path):
with open(path, "r") as f:
d = yaml.safe_load(f)
if d is None:
# no entries in the file
return
for name, spec in d.items():
assert name not in registry, f"duplicate entry: {name} from {path}"
if isinstance(spec, dict):
if "key" in spec:
raise ValueError(
f"key is a reserved keyword, but was used in {name} from {path}"
)
if "group" in spec:
raise ValueError(
f"group is a reserved keyword, but was used in {name} from {path}"
)
if "cls" in spec:
raise ValueError(
f"cls is a reserved keyword, but was used in {name} from {path}"
)
spec["key"] = name
spec["group"] = str(os.path.basename(path).split(".")[0])
if "class" in spec:
spec["cls"] = spec["class"]
del spec["class"]
registry[name] = spec
def _process_directory(self, registry, path):
files = Path(path).glob("*.yaml")
for file in files:
self._process_file(registry, file)
def _load_registry(self, paths):
"""Load registry from a list of paths.
Each path or yaml specifies a dictionary of name -> spec.
"""
registry = {}
for path in paths:
logging.info(f"Loading registry from {path}")
if os.path.exists(path):
if os.path.isdir(path):
self._process_directory(registry, path)
else:
self._process_file(registry, path)
return registry
@functools.cached_property
def _completion_fns(self):
return self._load_registry([p / "completion_fns" for p in self._registry_paths])
@functools.cached_property
def _eval_sets(self):
return self._load_registry([p / "eval_sets" for p in self._registry_paths])
@functools.cached_property
def _evals(self):
return self._load_registry([p / "evals" for p in self._registry_paths])
@functools.cached_property
def _modelgraded_specs(self):
return self._load_registry([p / "modelgraded" for p in self._registry_paths])
registry = Registry()