File size: 9,706 Bytes
df66f6e
2a5f9fb
 
df66f6e
2a5f9fb
 
 
 
 
20b060e
 
2a5f9fb
9d22eee
90fa47e
2a5f9fb
 
 
 
3dfaf22
 
 
 
2a5f9fb
3dfaf22
2a5f9fb
9d22eee
3dfaf22
9d22eee
3dfaf22
2a5f9fb
 
 
3dfaf22
2a5f9fb
90fa47e
 
2a5f9fb
 
 
3dfaf22
2a5f9fb
 
 
 
 
 
 
9d22eee
2a5f9fb
 
 
 
 
 
 
 
9d22eee
2a5f9fb
 
 
9d22eee
002172c
2a5f9fb
3dfaf22
9d22eee
002172c
3dfaf22
 
 
 
 
2a5f9fb
90fa47e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b554b5
90fa47e
 
 
 
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df66f6e
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
 
 
002172c
2a5f9fb
 
 
3dfaf22
 
2a5f9fb
20b060e
90fa47e
 
2a5f9fb
 
3dfaf22
 
9d22eee
2a5f9fb
 
 
 
 
9d22eee
2a5f9fb
 
 
b1a1395
2a5f9fb
 
 
 
3dfaf22
2a5f9fb
 
 
9d22eee
2a5f9fb
90fa47e
 
9d22eee
3dfaf22
2a5f9fb
 
 
 
 
 
 
 
90fa47e
 
2a5f9fb
 
 
 
 
 
 
 
3dfaf22
 
2a5f9fb
3dfaf22
2a5f9fb
 
 
 
 
 
 
 
 
 
 
9d22eee
2a5f9fb
 
 
 
 
 
3dfaf22
 
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
 
 
3dfaf22
2a5f9fb
 
3dfaf22
2a5f9fb
3dfaf22
 
2a5f9fb
 
 
 
 
 
 
 
 
 
 
b1a1395
 
df66f6e
2a5f9fb
 
 
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
import glob
import json
import math
import os
from dataclasses import dataclass

import dateutil
import numpy as np

from huggingface_hub import ModelCard

from src.display.formatting import make_clickable_model
from src.display.utils import AutoEvalColumn, ModelType, Tasks, Precision, WeightType
from src.submission.check_validity import is_model_on_hub, check_model_card


@dataclass
class EvalResult:
    # Also see src.display.utils.AutoEvalColumn for what will be displayed.
    eval_name: str # org_model_precision (uid)
    full_model: str # org/model (path on hub)
    org: str 
    model: str
    revision: str # commit hash, "" if main
    results: dict
    precision: Precision = Precision.Unknown
    model_type: ModelType = ModelType.Unknown # Pretrained, fine tuned, ...
    weight_type: WeightType = WeightType.Original # Original or Adapter
    architecture: str = "Unknown" # From config file
    license: str = "?"
    likes: int = 0
    num_params: int = 0
    date: str = "" # submission date of request file
    still_on_hub: bool = False
    is_merge: bool = False
    flagged: bool = False

    @classmethod
    def init_from_json_file(self, json_filepath):
        """Inits the result from the specific model result file"""
        with open(json_filepath) as fp:
            data = json.load(fp)

        # We manage the legacy config format
        config = data.get("config", data.get("config_general", None))

        # Precision
        precision = Precision.from_str(config.get("model_dtype"))

        # Get model and org
        org_and_model = config.get("model_name", config.get("model_args", None))
        org_and_model = org_and_model.split("/", 1)

        if len(org_and_model) == 1:
            org = None
            model = org_and_model[0]
            result_key = f"{model}_{precision.value.name}"
        else:
            org = org_and_model[0]
            model = org_and_model[1]
            result_key = f"{org}_{model}_{precision.value.name}"
        full_model = "/".join(org_and_model)

        still_on_hub, error, model_config = is_model_on_hub(
            full_model, config.get("model_sha", "main"), trust_remote_code=True, test_tokenizer=False
        )
        architecture = "?"
        if model_config is not None:
            architectures = getattr(model_config, "architectures", None)
            if architectures:
                architecture = ";".join(architectures)

        # If the model doesn't have a model card or a license, we consider it's deleted
        if still_on_hub:
            try:
                if check_model_card(full_model)[0] is False:
                    still_on_hub = False
            except Exception:
                still_on_hub = False

        # Check if the model is a merge
        is_merge_from_metadata = False
        flagged = False
        if still_on_hub:
            model_card = ModelCard.load(full_model)

            if model_card.data.tags:
                is_merge_from_metadata = "merge" in model_card.data.tags
            merge_keywords = ["mergekit", "merged model", "merge model", "merging"]
            # If the model is a merge but not saying it in the metadata, we flag it
            is_merge_from_model_card = any(keyword in model_card.text.lower() for keyword in merge_keywords)
            flagged = is_merge_from_model_card and not is_merge_from_metadata


        # Extract results available in this file (some results are split in several files)
        results = {}
        for task in Tasks:
            task = task.value
            # We skip old mmlu entries
            wrong_mmlu_version = False
            if task.benchmark == "hendrycksTest":
                for mmlu_k in ["harness|hendrycksTest-abstract_algebra|5", "hendrycksTest-abstract_algebra"]:
                    if mmlu_k in data["versions"] and data["versions"][mmlu_k] == 0:
                        wrong_mmlu_version = True

            if wrong_mmlu_version:
                continue

            # Some truthfulQA values are NaNs
            if task.benchmark == "truthfulqa:mc" and "harness|truthfulqa:mc|0" in data["results"]:
                if math.isnan(float(data["results"]["harness|truthfulqa:mc|0"][task.metric])):
                    results[task.benchmark] = 0.0
                    continue

            # We average all scores of a given metric (mostly for mmlu)
            accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark in k])
            if accs.size == 0 or any([acc is None for acc in accs]):
                continue

            mean_acc = np.mean(accs) * 100.0
            results[task.benchmark] = mean_acc

        return self(
            eval_name=result_key,
            full_model=full_model,
            org=org,
            model=model,
            results=results,
            precision=precision,  
            revision= config.get("model_sha", ""),
            still_on_hub=still_on_hub,
            architecture=architecture,
            is_merge=is_merge_from_metadata,
            flagged=flagged,
        )

    def update_with_request_file(self, requests_path):
        """Finds the relevant request file for the current model and updates info with it"""
        request_file = get_request_file_for_model(requests_path, self.full_model, self.precision.value.name)

        try:
            with open(request_file, "r") as f:
                request = json.load(f)
            self.model_type = ModelType.from_str(request.get("model_type", ""))
            self.weight_type = WeightType[request.get("weight_type", "Original")]
            self.license = request.get("license", "?")
            self.likes = request.get("likes", 0)
            self.num_params = request.get("params", 0)
            self.date = request.get("submitted_time", "")
        except Exception:
            print(f"Could not find request file for {self.org}/{self.model}")

    def to_dict(self):
        """Converts the Eval Result to a dict compatible with our dataframe display"""
        average = sum([v for v in self.results.values() if v is not None]) / len(Tasks)
        data_dict = {
            "eval_name": self.eval_name,  # not a column, just a save name,
            AutoEvalColumn.precision.name: self.precision.value.name,
            AutoEvalColumn.model_type.name: self.model_type.value.name,
            AutoEvalColumn.merged.name: self.is_merge,
            AutoEvalColumn.model_type_symbol.name: self.model_type.value.symbol, # + "🥦" if self.is_merge,
            AutoEvalColumn.weight_type.name: self.weight_type.value.name,
            AutoEvalColumn.architecture.name: self.architecture,
            AutoEvalColumn.model.name: make_clickable_model(self.full_model),
            AutoEvalColumn.dummy.name: self.full_model,
            AutoEvalColumn.revision.name: self.revision,
            AutoEvalColumn.average.name: average,
            AutoEvalColumn.license.name: self.license,
            AutoEvalColumn.likes.name: self.likes,
            AutoEvalColumn.params.name: self.num_params,
            AutoEvalColumn.still_on_hub.name: self.still_on_hub,
            AutoEvalColumn.flagged.name: self.flagged

        }

        for task in Tasks:
            data_dict[task.value.col_name] = self.results[task.value.benchmark]

        return data_dict


def get_request_file_for_model(requests_path, model_name, precision):
    """Selects the correct request file for a given model. Only keeps runs tagged as FINISHED"""
    request_files = os.path.join(
        requests_path,
        f"{model_name}_eval_request_*.json",
    )
    request_files = glob.glob(request_files)

    # Select correct request file (precision)
    request_file = ""
    request_files = sorted(request_files, reverse=True)
    for tmp_request_file in request_files:
        with open(tmp_request_file, "r") as f:
            req_content = json.load(f)
            if (
                req_content["status"] in ["FINISHED"]
                and req_content["precision"] == precision.split(".")[-1]
            ):
                request_file = tmp_request_file
    return request_file


def get_raw_eval_results(results_path: str, requests_path: str) -> list[EvalResult]:
    """From the path of the results folder root, extract all needed info for results"""
    model_result_filepaths = []

    for root, _, files in os.walk(results_path):
        # We should only have json files in model results
        if len(files) == 0 or any([not f.endswith(".json") for f in files]):
            continue

        # Sort the files by date
        try:
            files.sort(key=lambda x: x.removesuffix(".json").removeprefix("results_")[:-7])
        except dateutil.parser._parser.ParserError:
            files = [files[-1]]

        for file in files:
            model_result_filepaths.append(os.path.join(root, file))

    eval_results = {}
    for model_result_filepath in model_result_filepaths:
        # Creation of result
        eval_result = EvalResult.init_from_json_file(model_result_filepath)
        eval_result.update_with_request_file(requests_path)

        # Store results of same eval together
        eval_name = eval_result.eval_name
        if eval_name in eval_results.keys():
            eval_results[eval_name].results.update({k: v for k, v in eval_result.results.items() if v is not None})
        else:
            eval_results[eval_name] = eval_result

    results = []
    for v in eval_results.values():
        try:
            v.to_dict() # we test if the dict version is complete
            results.append(v)
        except KeyError:  # not all eval values present
            continue

    return results