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 | """Utility functions for the Every Eval Ever (EEE) output format."""
import collections.abc as c
import datetime
import json
import math
import time
import typing as t
from .constants import EEE_SCHEMA_VERSION
if t.TYPE_CHECKING:
from .data_models import BenchmarkResult
def benchmark_result_to_eee_dict(result: "BenchmarkResult") -> dict:
"""Convert a BenchmarkResult to the Every Eval Ever (EEE) format.
Produces a dictionary conforming to the Every Eval Ever JSON schema v0.2.1
(https://github.com/evaleval/every_eval_ever/blob/main/eval.schema.json).
The resulting dict can be written directly to
`euroeval_benchmark_results.jsonl` and later reconstructed without loss via
`benchmark_result_from_eee_dict`.
The mapping is as follows:
* Top-level fields: `schema_version`, `evaluation_id`,
`evaluation_timestamp`, `retrieved_timestamp`, `source_metadata`.
* `model_info`: model `id`/`name` plus EuroEval-specific details
(`num_model_parameters`, `max_sequence_length`, `vocabulary_size`,
`merge`, `generative`, `generative_type`) in `additional_details`.
* `eval_library`: `name="euroeval"`, library version, and evaluation
context (languages, task, shot config, library versions, raw per-iteration
scores) in `additional_details`.
* `evaluation_results`: one entry per metric. The 95 % confidence interval
half-width stored in the `_se` keys is exposed as a `confidence_interval`
with `confidence_level: 0.95`. Speed metrics (`test_speed`,
`test_speed_short`) do not include `score_type`, `min_score`, or
`max_score` because tokens-per-second has no fixed upper bound.
Args:
result:
The benchmark result to convert.
Returns:
A dictionary matching the EEE JSON schema v0.2.1.
"""
current_time = time.time()
retrieved_timestamp = str(int(current_time))
evaluation_timestamp = datetime.datetime.fromtimestamp(
current_time, tz=datetime.timezone.utc
).isoformat()
evaluation_id = f"{result.dataset}/{result.model}/{retrieved_timestamp}"
total_results: dict[str, float] = {}
raw_results: list = []
if isinstance(result.results, dict):
if "total" in result.results and isinstance(result.results["total"], dict):
total_results = result.results["total"]
if "raw" in result.results:
raw_results = list(result.results["raw"])
num_samples = len(raw_results)
metric_names = [
k
for k in total_results
if not k.endswith("_se") and k != "num_failed_instances"
]
num_failed = total_results.get("num_failed_instances", 0.0)
evaluation_results = []
for metric_name in metric_names:
score = total_results[metric_name]
ci_half_width = total_results.get(f"{metric_name}_se", float("nan"))
score_details: dict = {
"score": score,
"details": {"num_failed_instances": str(num_failed)},
}
uncertainty: dict = {}
if not math.isnan(ci_half_width):
uncertainty["confidence_interval"] = {
"lower": score - ci_half_width,
"upper": score + ci_half_width,
"confidence_level": 0.95,
"method": "bootstrap",
}
if num_samples > 0:
uncertainty["num_samples"] = num_samples
if uncertainty:
score_details["uncertainty"] = uncertainty
is_speed_metric = "speed" in metric_name
metric_config: dict = {"lower_is_better": False}
if not is_speed_metric:
metric_config["score_type"] = "continuous"
metric_config["min_score"] = 0
metric_config["max_score"] = 100
evaluation_results.append(
{
"evaluation_name": metric_name,
"source_data": {
"dataset_name": result.dataset,
"source_type": "hf_dataset",
},
"metric_config": metric_config,
"score_details": score_details,
}
)
inference_engine: dict = {}
if result.litellm_version:
inference_engine = {"name": "litellm", "version": result.litellm_version}
elif result.vllm_version:
inference_engine = {"name": "vllm", "version": result.vllm_version}
elif result.transformers_version:
inference_engine = {
"name": "transformers",
"version": result.transformers_version,
}
model_info: dict = {
"name": result.model,
"id": result.model,
"additional_details": {
"num_model_parameters": str(result.num_model_parameters),
"max_sequence_length": str(result.max_sequence_length),
"vocabulary_size": str(result.vocabulary_size),
"merge": str(result.merge).lower(),
"generative": str(result.generative).lower(),
"generative_type": result.generative_type
if result.generative_type is not None
else None,
},
}
if inference_engine:
model_info["inference_engine"] = inference_engine
eval_lib_additional_details = {
"dataset": result.dataset,
"languages": json.dumps(list(result.languages), ensure_ascii=False),
"task": result.task,
"few_shot": str(result.few_shot).lower()
if result.few_shot is not None
else None,
"validation_split": str(result.validation_split).lower()
if result.validation_split is not None
else None,
"transformers_version": result.transformers_version or None,
"torch_version": result.torch_version or None,
"vllm_version": result.vllm_version or None,
"xgrammar_version": result.xgrammar_version or None,
"litellm_version": result.litellm_version or None,
"raw_results": json.dumps(raw_results, ensure_ascii=False),
}
return {
"schema_version": EEE_SCHEMA_VERSION,
"evaluation_id": evaluation_id,
"evaluation_timestamp": evaluation_timestamp,
"retrieved_timestamp": retrieved_timestamp,
"source_metadata": {
"source_name": "EuroEval",
"source_type": "evaluation_run",
"source_organization_name": "EuroEval",
"source_organization_url": "https://euroeval.com",
"evaluator_relationship": "third_party",
},
"model_info": model_info,
"eval_library": {
"name": "euroeval",
"version": result.euroeval_version or "unknown",
"additional_details": eval_lib_additional_details,
},
"evaluation_results": evaluation_results,
}
def benchmark_result_from_eee_dict(config: dict) -> "BenchmarkResult":
"""Create a BenchmarkResult from an Every Eval Ever format dictionary.
Reconstructs a full `BenchmarkResult` from a dictionary conforming to the
Every Eval Ever (EEE) JSON schema v0.2.1. This function is the inverse of
`benchmark_result_to_eee_dict` and enables lossless round-trips.
Args:
config:
A dictionary conforming to the EEE JSON schema v0.2.1, as produced
by `benchmark_result_to_eee_dict`.
Returns:
The reconstructed benchmark result.
"""
# Importing here to avoid circular imports
from .data_models import BenchmarkResult # noqa: PLC0415
model_info = config.get("model_info", {})
eval_library = config.get("eval_library", {})
evaluation_results: c.Sequence[dict] = config.get("evaluation_results", [])
model = model_info.get("id", "")
model_additional = model_info.get("additional_details", {})
eval_lib_additional = eval_library.get("additional_details", {})
if evaluation_results:
dataset = evaluation_results[0].get("source_data", {}).get("dataset_name", "")
else:
dataset = eval_lib_additional.get("dataset", "")
raw_results_json = eval_lib_additional.get("raw_results", "[]")
try:
raw_results = json.loads(raw_results_json)
except json.JSONDecodeError:
raw_results = []
total_dict: dict[str, float] = {}
for eval_result in evaluation_results:
metric_name = eval_result.get("evaluation_name", "")
score_details = eval_result.get("score_details", {})
total_dict[metric_name] = score_details.get("score", 0.0)
uncertainty = score_details.get("uncertainty", {})
ci_info = uncertainty.get("confidence_interval", {})
if "lower" in ci_info and "upper" in ci_info:
ci_half_width = (ci_info["upper"] - ci_info["lower"]) / 2
total_dict[f"{metric_name}_se"] = ci_half_width
details = score_details.get("details", {})
if (
"num_failed_instances" in details
and "num_failed_instances" not in total_dict
):
total_dict["num_failed_instances"] = float(details["num_failed_instances"])
results = {"raw": raw_results, "total": total_dict}
languages_json = eval_lib_additional.get("languages", "[]")
try:
languages: list[str] = json.loads(languages_json)
except json.JSONDecodeError:
languages = []
return BenchmarkResult(
dataset=dataset,
task=eval_lib_additional.get("task", ""),
languages=languages,
model=model,
results=results,
num_model_parameters=int(
model_additional.get("num_model_parameters", "0") or "0"
),
max_sequence_length=int(
model_additional.get("max_sequence_length", "0") or "0"
),
vocabulary_size=int(model_additional.get("vocabulary_size", "0") or "0"),
merge=model_additional.get("merge", "false") == "true",
generative=model_additional.get("generative", "false") == "true",
generative_type=parse_optional_str(model_additional.get("generative_type")),
few_shot=parse_optional_bool(eval_lib_additional.get("few_shot")),
validation_split=parse_optional_bool(
eval_lib_additional.get("validation_split")
),
euroeval_version=parse_optional_str(
None
if eval_library.get("version") == "unknown"
else eval_library.get("version")
),
transformers_version=parse_optional_str(
eval_lib_additional.get("transformers_version")
),
torch_version=parse_optional_str(eval_lib_additional.get("torch_version")),
vllm_version=parse_optional_str(eval_lib_additional.get("vllm_version")),
xgrammar_version=parse_optional_str(
eval_lib_additional.get("xgrammar_version")
),
litellm_version=parse_optional_str(eval_lib_additional.get("litellm_version")),
)
def parse_optional_str(value: str | None) -> str | None:
"""Parse a string-encoded optional string value.
Args:
value:
The string to parse. `None` maps to `None`.
Returns:
`None` if value is `None`, otherwise the original string.
"""
return None if value is None else value
def parse_optional_bool(value: str | None) -> bool | None:
"""Parse a string-encoded optional boolean value.
Args:
value:
The string to parse. `None` maps to `None`; any other value is
compared case-insensitively to `"true"`.
Returns:
`None` if value is `None`, otherwise a boolean.
"""
if value is None:
return None
return value.lower() == "true"
|