forked from openai/evals
-
Notifications
You must be signed in to change notification settings - Fork 3
/
classify.py
314 lines (278 loc) · 13.9 KB
/
classify.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
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
"""
Generic eval that uses a prompt + classification.
"""
import itertools
import string
from collections import Counter
from random import Random
from typing import Callable, Iterable, Optional
import openai
import evals
import evals.record
from evals.base import ModelSpec
from evals.elsuite.utils import (
PromptFn,
format_necessary,
load_modelgraded_specs,
scrub_formatting_from_prompt,
)
INVALID_STR = "__invalid__"
CHOICE_KEY = "choice"
MATCH_FNS = {
"include": lambda x, y: float(x in y),
"exact": lambda x, y: float(x == y),
"endswith": lambda x, y: x.endswith(y),
}
CHOICE_FNS = {
# e.g. "Yes"
"classify": lambda x: x.strip(),
# e.g. "Yes\n The reasons are: ..."
"classify_cot": lambda x: x.strip().split("\n")[0].strip(),
# e.g. "Let's think step by step. ...\nYes"
"cot_classify": lambda x: x.strip().split("\n")[-1].strip(),
}
ANSWER_PROMPTS = {
"classify": "Answer the question by printing only a single choice from {choices} (without quotes or punctuation) corresponding to the correct answer with no other text.".strip(),
"classify_cot": "First, answer by printing a single choice from {choices} (without quotes or punctuation) corresponding to the correct answer. Then, from the next line, explain your reasonings step by step.".strip(),
"cot_classify": """
First, write out in a step by step manner your reasoning to be sure that your conclusion is correct. Avoid simply stating the correct answer at the outset. Then print only a single choice from {choices} (without quotes or punctuation) on its own line corresponding to the correct answer. At the end, repeat just the answer by itself on a new line.
Reasoning:""".strip(),
}
EVAL_MODELSPEC = ModelSpec(name="gpt-3.5-turbo", model="gpt-3.5-turbo", is_chat=True)
def choice_to_str(choice_strings: Iterable[str]) -> str:
"""Return a string of choices, e.g. '"Yes" or "No" or "Maybe"'."""
return " or ".join(f'"{choice}"' for choice in choice_strings)
def clean_choice(raw_choice: str, match_fn: Callable, choice_strings: Iterable[str]) -> str:
"""Clean a choice string to one of choice_strings. Return '__invalid__.' if no match."""
raw_choice = raw_choice.strip()
raw_choice = "".join(c for c in raw_choice if c not in string.punctuation)
if not raw_choice:
return INVALID_STR
for choice in choice_strings:
if match_fn(raw_choice, choice):
return choice
return INVALID_STR
def expand_args_dict(args_dict):
"""Expand a dict of dicts, with namings.
args_dict = {
"a": {"a1": 1, "a2": 2},
"b": {"b1": 3, "b2": 4},
}
expand_args_dict(args_dict) = {
"a=a1:b=b1": {"a": ("a1", 1), "b": ("b1", 3)},
"a=a1:b=b2": {"a": ("a1", 1), "b": ("b2", 4)},
...}
"""
args_dict = {k: list(v.items()) for k, v in args_dict.items()}
keys = list(args_dict.keys())
values = list(args_dict.values())
new_values = [dict(zip(keys, v)) for v in itertools.product(*values)]
new_names = [":".join([f"{k}={v[0]}" for k, v in sorted(d.items())]) for d in new_values]
return dict(zip(new_names, new_values))
class ModelBasedClassify(evals.Eval):
invalid_request_during_completion = 0
invalid_request_during_evaluation = 0
def __init__(
self,
model_specs: evals.ModelSpecs,
samples_jsonl: str,
modelgraded_spec_file: str,
*args,
match_fn: str = "endswith",
max_tokens: int = 1024,
multicomp_n: int = 1,
multicomp_temperature: float = 0.4,
samples_renamings: Optional[dict[str, str]] = None,
eval_type: Optional[str] = None,
metaeval: bool = False,
**kwargs,
):
super().__init__(model_specs, *args, **kwargs)
self.max_tokens = max_tokens
self.samples_jsonl = samples_jsonl
self.match_fn = MATCH_FNS[match_fn]
self.metaeval = metaeval
self.multicomp_n = multicomp_n
self.multicomp_temperature = multicomp_temperature
self.samples_renamings = samples_renamings or {}
"""import prompt and set attributes"""
modelgraded_specs = load_modelgraded_specs(modelgraded_spec_file)
# 'choice_strings' is a list of strings that specifies the possible choices
self.choice_strings = modelgraded_specs.pop("choice_strings")
# make sure each choice doesn't contain any punctuation
for s in self.choice_strings:
assert not any(c in s for c in string.punctuation), f"{s} contains punctuation"
# (optional) 'choice_scores' is a dict that specifies the score for each choice string
# if 'choice_scores' is specified, 'scores/' are computed and added to metrics
self.choice_scores = modelgraded_specs.pop("choice_scores", {})
if self.choice_scores == "from_strings":
self.choice_scores = {c: float(c) for c in self.choice_strings}
assert all(
isinstance(v, (int, float)) for v in self.choice_scores.values()
), f"choice_scores must be a dict of floats, not {self.choice_scores}"
# (optional) 'eval_type' is a string that specifies the type of classification algorithm
# - "classify": only answer
# - "cot_classify": reason then answer (chain-of-thought) <- most recommended
# - "classify_cot": answer then reason (explanation)
# if 'eval_type' is not supplied from modelgraded_specs, then it must be supplied as an argument.
# - Importantly, it also assumes the answer prompt needs to be appended to the prompt.
# 'eval_type' sets 'choice_fn', a function that takes the model's raw response and returns the choice string
self.eval_type = modelgraded_specs.pop("eval_type", None)
if not self.eval_type:
append_answer_prompt = True # append answer prompt to prompt
assert (
eval_type
), "eval_type must be specified, in modelgraded_spec_file or as an argument"
self.eval_type = eval_type
else:
assert (
not eval_type
), f"eval_type must be unspecified, if it is specified in modelgraded_spec_file"
append_answer_prompt = False
assert self.eval_type in CHOICE_FNS, f"eval_type must be one of {list(CHOICE_FNS.keys())}"
self.choice_fn = CHOICE_FNS[self.eval_type]
# 'prompt' is a string that specifies the model-graded evaluation
prompt = modelgraded_specs.pop("prompt")
assert isinstance(prompt, str), f"prompt must be a string, not {type(prompt)}"
if append_answer_prompt:
prompt += "\n\n" + ANSWER_PROMPTS[self.eval_type].format(
choices=choice_to_str(self.choice_strings)
)
self.prompt = [{"role": "user", "content": prompt}]
# 'input_outputs' is a dict that specifies the input and output keys in the sample
# output key is the model's raw response to input key. These are used for filling 'prompt' template.
self.input_outputs = modelgraded_specs.pop("input_outputs")
assert isinstance(
self.input_outputs, dict
), f"input_outputs must be a dict, not {type(self.input_outputs)}"
# (optional) 'args' is a dict of dicts that specifies additional arguments for 'prompt'
# each value in 'args_dict' essentially defines a separate modelgraded classification eval and has own metrics!
self.args_dict = modelgraded_specs.pop("args", {})
if self.args_dict:
self.expanded_args_dict = expand_args_dict(self.args_dict)
else:
self.expanded_args_dict = {}
# (optional) 'completion_sample_templates'
# each key must be one of 'input_outputs'.values(). If 'multicomp_n' > 1, this template is filled 'multicomp_n' times
# and the concatenated result is passed to 'prompt' template.
self.completion_sample_templates = modelgraded_specs.pop("completion_sample_templates", {})
assert all(
k in self.input_outputs.values() for k in self.completion_sample_templates
), f"all {self.completion_sample_templates.keys()} must be in {self.input_outputs.values()}, "
if self.multicomp_n > 1:
assert (
self.completion_sample_templates
), "completion_sample_templates must be specified if multicomp_n > 1"
# since we accept optional args, we need to check that all args are used
assert not modelgraded_specs, f"Unused args: {modelgraded_specs}. Typo in YAML?"
def eval_sample(self, test_sample: dict, rng: Random) -> None:
"""Evaluate a single sample.
Recorded metrics are always: one of the self.choice_strings, or "__invalid__".
"""
if self.samples_renamings:
test_sample = {self.samples_renamings.get(k, k): v for k, v in test_sample.items()}
if self.multicomp_n > 1:
test_sample["n"] = self.multicomp_n
completions = {}
if self.metaeval:
# assert outputs exist in the data
for v in self.input_outputs.values():
assert v in test_sample, f"Missing output '{v}' in sample {test_sample.keys()}"
completions[v] = test_sample[v]
# remove outputs from the data
test_sample = {
k: v for k, v in test_sample.items() if k not in list(self.input_outputs.values())
}
for k in self.input_outputs:
test_sample[k] = scrub_formatting_from_prompt(test_sample[k])
if not self.metaeval:
try:
for k, v in self.input_outputs.items():
if self.multicomp_n > 1 and v in self.completion_sample_templates:
completion = ""
completion_i_template = self.completion_sample_templates[v]
for i in range(self.multicomp_n):
get_input_completion = PromptFn(
test_sample[k],
model_spec=self.model_spec,
max_tokens=self.max_tokens,
temperature=self.multicomp_temperature,
)
completion_i, _ = get_input_completion()
completion += format_necessary(
completion_i_template,
i=i + 1,
output=completion_i,
n=self.multicomp_n,
)
else:
get_input_completion = PromptFn(
test_sample[k],
model_spec=self.model_spec,
max_tokens=self.max_tokens,
)
completion, _ = get_input_completion()
completions[v] = completion
except openai.error.InvalidRequestError:
self.invalid_request_during_completion += 1
return
try:
metrics = {}
evaluate = PromptFn(
self.prompt,
model_spec=EVAL_MODELSPEC,
max_tokens=self.max_tokens,
)
eval_kwargs = dict(**completions, **test_sample)
if self.expanded_args_dict:
args_dict = self.expanded_args_dict
else:
args_dict = {CHOICE_KEY: {}}
for metric, args in args_dict.items():
args = {k: v[1] for k, v in args.items()}
evaluation, _ = evaluate(**args, **eval_kwargs)
raw_choice = self.choice_fn(evaluation)
choice = clean_choice(raw_choice, self.match_fn, self.choice_strings)
metrics[metric] = choice
if self.metaeval:
assert (
metric in test_sample
), f"Missing label for metric '{metric}' in sample {test_sample.keys()}"
metrics[metric + "_metascore"] = choice == test_sample[metric]
except openai.error.InvalidRequestError:
self.invalid_request_during_evaluation += 1
return
evals.record.record_metrics(**metrics)
return choice
def run(self, recorder):
samples = evals.get_jsonl(self.samples_jsonl)
self.eval_all_samples(recorder, samples)
all_sample_metrics = recorder.get_metrics()
record_metrics = {}
if self.expanded_args_dict:
metrics = sorted(self.expanded_args_dict)
else:
metrics = [CHOICE_KEY]
for metric in metrics:
chosen = [m[metric] for m in all_sample_metrics if metric in m]
# if there is a best choice, compute the score
if self.choice_scores:
# assumption: each INVALID_STR contributes the lowest score
lowest_score = min(self.choice_scores.values())
scores = [
self.choice_scores[choice] if choice != INVALID_STR else lowest_score
for choice in chosen
]
record_metrics[f"score/{metric}"] = sum(scores) / len(all_sample_metrics)
# compute the counts and ratios
counts = dict(Counter(chosen))
missing_samples = len(all_sample_metrics) - len(chosen)
if missing_samples:
counts["__missing_samples__"] = missing_samples
record_metrics.update({f"counts/{metric}/{k}": v for k, v in counts.items()})
if self.metaeval:
metascores = [m[metric + "_metascore"] for m in all_sample_metrics if metric in m]
record_metrics[f"metascore/{metric}"] = sum(metascores) / len(all_sample_metrics)
record_metrics["invalid_request_during_completion"] = self.invalid_request_during_completion
record_metrics["invalid_request_during_evaluation"] = self.invalid_request_during_evaluation
return record_metrics