nemo-evaluator-launcher 0.1.0rc2__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Potentially problematic release.
This version of nemo-evaluator-launcher might be problematic. Click here for more details.
- nemo_evaluator_launcher/__init__.py +65 -0
- nemo_evaluator_launcher/api/__init__.py +24 -0
- nemo_evaluator_launcher/api/functional.py +641 -0
- nemo_evaluator_launcher/api/types.py +89 -0
- nemo_evaluator_launcher/api/utils.py +19 -0
- nemo_evaluator_launcher/cli/__init__.py +15 -0
- nemo_evaluator_launcher/cli/export.py +148 -0
- nemo_evaluator_launcher/cli/info.py +117 -0
- nemo_evaluator_launcher/cli/kill.py +39 -0
- nemo_evaluator_launcher/cli/ls_runs.py +113 -0
- nemo_evaluator_launcher/cli/ls_tasks.py +34 -0
- nemo_evaluator_launcher/cli/main.py +136 -0
- nemo_evaluator_launcher/cli/run.py +135 -0
- nemo_evaluator_launcher/cli/status.py +118 -0
- nemo_evaluator_launcher/cli/version.py +52 -0
- nemo_evaluator_launcher/common/__init__.py +16 -0
- nemo_evaluator_launcher/common/execdb.py +189 -0
- nemo_evaluator_launcher/common/helpers.py +157 -0
- nemo_evaluator_launcher/common/logging_utils.py +349 -0
- nemo_evaluator_launcher/common/mapping.py +310 -0
- nemo_evaluator_launcher/configs/__init__.py +15 -0
- nemo_evaluator_launcher/configs/default.yaml +28 -0
- nemo_evaluator_launcher/configs/deployment/nim.yaml +32 -0
- nemo_evaluator_launcher/configs/deployment/none.yaml +16 -0
- nemo_evaluator_launcher/configs/deployment/sglang.yaml +38 -0
- nemo_evaluator_launcher/configs/deployment/vllm.yaml +41 -0
- nemo_evaluator_launcher/configs/execution/lepton/default.yaml +92 -0
- nemo_evaluator_launcher/configs/execution/local.yaml +17 -0
- nemo_evaluator_launcher/configs/execution/slurm/default.yaml +33 -0
- nemo_evaluator_launcher/executors/__init__.py +22 -0
- nemo_evaluator_launcher/executors/base.py +97 -0
- nemo_evaluator_launcher/executors/lepton/__init__.py +16 -0
- nemo_evaluator_launcher/executors/lepton/deployment_helpers.py +589 -0
- nemo_evaluator_launcher/executors/lepton/executor.py +905 -0
- nemo_evaluator_launcher/executors/lepton/job_helpers.py +394 -0
- nemo_evaluator_launcher/executors/local/__init__.py +15 -0
- nemo_evaluator_launcher/executors/local/executor.py +491 -0
- nemo_evaluator_launcher/executors/local/run.template.sh +88 -0
- nemo_evaluator_launcher/executors/registry.py +38 -0
- nemo_evaluator_launcher/executors/slurm/__init__.py +15 -0
- nemo_evaluator_launcher/executors/slurm/executor.py +982 -0
- nemo_evaluator_launcher/exporters/__init__.py +36 -0
- nemo_evaluator_launcher/exporters/base.py +112 -0
- nemo_evaluator_launcher/exporters/gsheets.py +391 -0
- nemo_evaluator_launcher/exporters/local.py +488 -0
- nemo_evaluator_launcher/exporters/mlflow.py +448 -0
- nemo_evaluator_launcher/exporters/registry.py +40 -0
- nemo_evaluator_launcher/exporters/utils.py +669 -0
- nemo_evaluator_launcher/exporters/wandb.py +376 -0
- nemo_evaluator_launcher/package_info.py +35 -0
- nemo_evaluator_launcher/resources/mapping.toml +344 -0
- nemo_evaluator_launcher-0.1.0rc2.dist-info/METADATA +35 -0
- nemo_evaluator_launcher-0.1.0rc2.dist-info/RECORD +57 -0
- nemo_evaluator_launcher-0.1.0rc2.dist-info/WHEEL +5 -0
- nemo_evaluator_launcher-0.1.0rc2.dist-info/entry_points.txt +3 -0
- nemo_evaluator_launcher-0.1.0rc2.dist-info/licenses/LICENSE +451 -0
- nemo_evaluator_launcher-0.1.0rc2.dist-info/top_level.txt +1 -0
|
@@ -0,0 +1,344 @@
|
|
|
1
|
+
# NOTE(agronskiy): checked parity
|
|
2
|
+
[lm-evaluation-harness]
|
|
3
|
+
container = "nvcr.io/nvidia/eval-factory/lm-evaluation-harness:25.08.1"
|
|
4
|
+
|
|
5
|
+
[lm-evaluation-harness.tasks.chat.ifeval]
|
|
6
|
+
required_env_vars = []
|
|
7
|
+
|
|
8
|
+
[lm-evaluation-harness.tasks.chat.mmlu_prox]
|
|
9
|
+
required_env_vars = []
|
|
10
|
+
|
|
11
|
+
[lm-evaluation-harness.tasks.completions.mmlu]
|
|
12
|
+
required_env_vars = []
|
|
13
|
+
|
|
14
|
+
[lm-evaluation-harness.tasks.completions.mmlu_pro]
|
|
15
|
+
|
|
16
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu]
|
|
17
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_ar]
|
|
18
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_bn]
|
|
19
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_de]
|
|
20
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_en]
|
|
21
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_es]
|
|
22
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_fr]
|
|
23
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_hi]
|
|
24
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_id]
|
|
25
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_it]
|
|
26
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_ja]
|
|
27
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_ko]
|
|
28
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_pt]
|
|
29
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_sw]
|
|
30
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_yo]
|
|
31
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_zh]
|
|
32
|
+
|
|
33
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full]
|
|
34
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_am]
|
|
35
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ar]
|
|
36
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_bn]
|
|
37
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_cs]
|
|
38
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_de]
|
|
39
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_el]
|
|
40
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_en]
|
|
41
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_es]
|
|
42
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_fa]
|
|
43
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_fil]
|
|
44
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_fr]
|
|
45
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ha]
|
|
46
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_he]
|
|
47
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_hi]
|
|
48
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_id]
|
|
49
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ig]
|
|
50
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_it]
|
|
51
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ja]
|
|
52
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ko]
|
|
53
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ky]
|
|
54
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_lt]
|
|
55
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_mg]
|
|
56
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ms]
|
|
57
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ne]
|
|
58
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_nl]
|
|
59
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ny]
|
|
60
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_pl]
|
|
61
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_pt]
|
|
62
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ro]
|
|
63
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_ru]
|
|
64
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_si]
|
|
65
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_sn]
|
|
66
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_so]
|
|
67
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_sr]
|
|
68
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_sv]
|
|
69
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_sw]
|
|
70
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_te]
|
|
71
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_tr]
|
|
72
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_uk]
|
|
73
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_vi]
|
|
74
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_yo]
|
|
75
|
+
[lm-evaluation-harness.tasks.completions.global_mmlu_full_zh]
|
|
76
|
+
[lm-evaluation-harness.tasks.completions.mmlu_logits]
|
|
77
|
+
|
|
78
|
+
[lm-evaluation-harness.tasks.chat.mmlu_instruct]
|
|
79
|
+
|
|
80
|
+
[lm-evaluation-harness.tasks.chat.mmlu_redux_instruct]
|
|
81
|
+
|
|
82
|
+
[lm-evaluation-harness.tasks.completions.gsm8k]
|
|
83
|
+
required_env_vars = []
|
|
84
|
+
|
|
85
|
+
[lm-evaluation-harness.tasks.chat.gsm8k_cot_instruct]
|
|
86
|
+
required_env_vars = []
|
|
87
|
+
|
|
88
|
+
[lm-evaluation-harness.tasks.chat.gsm8k_cot_llama]
|
|
89
|
+
required_env_vars = []
|
|
90
|
+
|
|
91
|
+
[lm-evaluation-harness.tasks.chat.mgsm_cot]
|
|
92
|
+
|
|
93
|
+
[lm-evaluation-harness.tasks.chat.gpqa_diamond_cot]
|
|
94
|
+
|
|
95
|
+
[lm-evaluation-harness.tasks.completions.winogrande]
|
|
96
|
+
|
|
97
|
+
[lm-evaluation-harness.tasks.completions.hellaswag]
|
|
98
|
+
[lm-evaluation-harness.tasks.completions.hellaswag_multilingual]
|
|
99
|
+
|
|
100
|
+
[lm-evaluation-harness.tasks.completions.commonsense_qa]
|
|
101
|
+
|
|
102
|
+
[lm-evaluation-harness.tasks.completions.openbookqa]
|
|
103
|
+
|
|
104
|
+
[lm-evaluation-harness.tasks.completions.piqa]
|
|
105
|
+
|
|
106
|
+
[lm-evaluation-harness.tasks.completions.adlr_race]
|
|
107
|
+
|
|
108
|
+
[lm-evaluation-harness.tasks.completions.social_iqa]
|
|
109
|
+
|
|
110
|
+
[lm-evaluation-harness.tasks.completions.adlr_truthfulqa_mc2]
|
|
111
|
+
[lm-evaluation-harness.tasks.completions.adlr_minerva_math_nemo]
|
|
112
|
+
[lm-evaluation-harness.tasks.completions.adlr_arc_challenge_llama]
|
|
113
|
+
[lm-evaluation-harness.tasks.completions.adlr_mmlu_pro_5_shot_base]
|
|
114
|
+
[lm-evaluation-harness.tasks.completions.adlr_mbpp_sanitized_3shot_greedy]
|
|
115
|
+
[lm-evaluation-harness.tasks.completions.adlr_mbppplus_greedy_sanitized]
|
|
116
|
+
[lm-evaluation-harness.tasks.completions.adlr_humaneval_greedy]
|
|
117
|
+
[lm-evaluation-harness.tasks.completions.adlr_humanevalplus_greedy]
|
|
118
|
+
[lm-evaluation-harness.tasks.chat.adlr_gsm8k_fewshot_cot]
|
|
119
|
+
required_env_vars = []
|
|
120
|
+
|
|
121
|
+
[lm-evaluation-harness.tasks.completions.arc_multilingual]
|
|
122
|
+
|
|
123
|
+
|
|
124
|
+
###############################################################################
|
|
125
|
+
# NOTE(agronskiy): checked parity
|
|
126
|
+
[mtbench]
|
|
127
|
+
container = "nvcr.io/nvidia/eval-factory/mtbench:25.08.1"
|
|
128
|
+
|
|
129
|
+
[mtbench.tasks.chat.mtbench]
|
|
130
|
+
|
|
131
|
+
[mtbench.tasks.chat.mtbench-cor1]
|
|
132
|
+
|
|
133
|
+
|
|
134
|
+
###############################################################################
|
|
135
|
+
# NOTE(agronskiy): checked parity
|
|
136
|
+
[ifbench]
|
|
137
|
+
container = "nvcr.io/nvidia/eval-factory/ifbench:25.08.1"
|
|
138
|
+
|
|
139
|
+
[ifbench.tasks.chat.ifbench]
|
|
140
|
+
required_env_vars = []
|
|
141
|
+
|
|
142
|
+
|
|
143
|
+
###############################################################################
|
|
144
|
+
[simple_evals]
|
|
145
|
+
container = "nvcr.io/nvidia/eval-factory/simple-evals:25.08.1"
|
|
146
|
+
|
|
147
|
+
[simple_evals.tasks.chat.gpqa_diamond]
|
|
148
|
+
required_env_vars = ["HF_TOKEN"]
|
|
149
|
+
|
|
150
|
+
[simple_evals.tasks.chat.gpqa_diamond_aa_v2]
|
|
151
|
+
required_env_vars = ["HF_TOKEN"]
|
|
152
|
+
|
|
153
|
+
[simple_evals.tasks.chat.gpqa_diamond_aa_v2_llama_4]
|
|
154
|
+
required_env_vars = ["HF_TOKEN"]
|
|
155
|
+
|
|
156
|
+
[simple_evals.tasks.chat.gpqa_diamond_nemo]
|
|
157
|
+
required_env_vars = ["HF_TOKEN"]
|
|
158
|
+
|
|
159
|
+
[simple_evals.tasks.chat.AA_math_test_500]
|
|
160
|
+
required_env_vars = ["JUDGE_API_KEY"]
|
|
161
|
+
|
|
162
|
+
[simple_evals.tasks.chat.math_test_500_nemo]
|
|
163
|
+
required_env_vars = []
|
|
164
|
+
|
|
165
|
+
[simple_evals.tasks.chat.aime_2024_nemo]
|
|
166
|
+
required_env_vars = []
|
|
167
|
+
|
|
168
|
+
[simple_evals.tasks.chat.AA_AIME_2024]
|
|
169
|
+
required_env_vars = ["JUDGE_API_KEY"]
|
|
170
|
+
|
|
171
|
+
[simple_evals.tasks.chat.aime_2025_nemo]
|
|
172
|
+
required_env_vars = []
|
|
173
|
+
|
|
174
|
+
[simple_evals.tasks.chat.AIME_2025]
|
|
175
|
+
required_env_vars = ["JUDGE_API_KEY"]
|
|
176
|
+
|
|
177
|
+
[simple_evals.tasks.chat.humaneval]
|
|
178
|
+
required_env_vars = []
|
|
179
|
+
|
|
180
|
+
[simple_evals.tasks.chat.mgsm]
|
|
181
|
+
required_env_vars = []
|
|
182
|
+
|
|
183
|
+
[simple_evals.tasks.chat.mmlu_pro]
|
|
184
|
+
required_env_vars = []
|
|
185
|
+
|
|
186
|
+
[simple_evals.tasks.chat.mmlu]
|
|
187
|
+
required_env_vars = []
|
|
188
|
+
|
|
189
|
+
[simple_evals.tasks.chat.mmlu_llama_4]
|
|
190
|
+
required_env_vars = []
|
|
191
|
+
|
|
192
|
+
[simple_evals.tasks.chat.mmlu_pro_llama_4]
|
|
193
|
+
required_env_vars = []
|
|
194
|
+
|
|
195
|
+
[simple_evals.tasks.chat.mmlu_ar-lite]
|
|
196
|
+
[simple_evals.tasks.chat.mmlu_bn-lite]
|
|
197
|
+
[simple_evals.tasks.chat.mmlu_de-lite]
|
|
198
|
+
[simple_evals.tasks.chat.mmlu_en-lite]
|
|
199
|
+
[simple_evals.tasks.chat.mmlu_es-lite]
|
|
200
|
+
[simple_evals.tasks.chat.mmlu_fr-lite]
|
|
201
|
+
[simple_evals.tasks.chat.mmlu_hi-lite]
|
|
202
|
+
[simple_evals.tasks.chat.mmlu_id-lite]
|
|
203
|
+
[simple_evals.tasks.chat.mmlu_it-lite]
|
|
204
|
+
[simple_evals.tasks.chat.mmlu_ja-lite]
|
|
205
|
+
[simple_evals.tasks.chat.mmlu_ko-lite]
|
|
206
|
+
[simple_evals.tasks.chat.mmlu_my-lite]
|
|
207
|
+
[simple_evals.tasks.chat.mmlu_pt-lite]
|
|
208
|
+
[simple_evals.tasks.chat.mmlu_sw-lite]
|
|
209
|
+
[simple_evals.tasks.chat.mmlu_yo-lite]
|
|
210
|
+
[simple_evals.tasks.chat.mmlu_zh-lite]
|
|
211
|
+
|
|
212
|
+
|
|
213
|
+
###############################################################################
|
|
214
|
+
# NOTE(agronskiy): checked parity
|
|
215
|
+
[bigcode-evaluation-harness]
|
|
216
|
+
container = "nvcr.io/nvidia/eval-factory/bigcode-evaluation-harness:25.08.1"
|
|
217
|
+
|
|
218
|
+
[bigcode-evaluation-harness.tasks.chat.mbpp]
|
|
219
|
+
required_env_vars = []
|
|
220
|
+
|
|
221
|
+
[bigcode-evaluation-harness.tasks.chat.mbppplus]
|
|
222
|
+
|
|
223
|
+
[bigcode-evaluation-harness.tasks.chat.mbppplus_nemo]
|
|
224
|
+
required_env_vars = []
|
|
225
|
+
|
|
226
|
+
[bigcode-evaluation-harness.tasks.completions.humaneval]
|
|
227
|
+
required_env_vars = []
|
|
228
|
+
|
|
229
|
+
[bigcode-evaluation-harness.tasks.completions.humaneval_instruct]
|
|
230
|
+
|
|
231
|
+
|
|
232
|
+
###############################################################################
|
|
233
|
+
[livecodebench]
|
|
234
|
+
container = "nvcr.io/nvidia/eval-factory/livecodebench:25.08.1"
|
|
235
|
+
|
|
236
|
+
[livecodebench.tasks.chat.livecodebench_0724_0125]
|
|
237
|
+
required_env_vars = []
|
|
238
|
+
|
|
239
|
+
[livecodebench.tasks.chat.livecodebench_0824_0225]
|
|
240
|
+
required_env_vars = []
|
|
241
|
+
|
|
242
|
+
|
|
243
|
+
###############################################################################
|
|
244
|
+
[scicode]
|
|
245
|
+
container = "nvcr.io/nvidia/eval-factory/scicode:25.08.1"
|
|
246
|
+
|
|
247
|
+
[scicode.tasks.chat.aa_scicode]
|
|
248
|
+
required_env_vars = []
|
|
249
|
+
|
|
250
|
+
|
|
251
|
+
###############################################################################
|
|
252
|
+
[hle]
|
|
253
|
+
container = "nvcr.io/nvidia/eval-factory/hle:25.08.1"
|
|
254
|
+
|
|
255
|
+
[hle.tasks.chat.hle]
|
|
256
|
+
required_env_vars = ["HF_TOKEN", "OPENAI_CLIENT_ID", "OPENAI_CLIENT_SECRET"]
|
|
257
|
+
|
|
258
|
+
|
|
259
|
+
###############################################################################
|
|
260
|
+
[bfcl]
|
|
261
|
+
container = "nvcr.io/nvidia/eval-factory/bfcl:25.08.1"
|
|
262
|
+
|
|
263
|
+
[bfcl.tasks.chat.bfclv2_ast_prompting]
|
|
264
|
+
required_env_vars = []
|
|
265
|
+
|
|
266
|
+
[bfcl.tasks.chat.bfclv3_ast_prompting]
|
|
267
|
+
required_env_vars = []
|
|
268
|
+
|
|
269
|
+
|
|
270
|
+
###############################################################################
|
|
271
|
+
[vlmevalkit]
|
|
272
|
+
container = "nvcr.io/nvidia/eval-factory/vlmevalkit:25.08.1"
|
|
273
|
+
|
|
274
|
+
[vlmevalkit.tasks.vlm.ocrbench]
|
|
275
|
+
required_env_vars = []
|
|
276
|
+
|
|
277
|
+
[vlmevalkit.tasks.vlm.slidevqa]
|
|
278
|
+
required_env_vars = ["OPENAI_CLIENT_ID", "OPENAI_CLIENT_SECRET"]
|
|
279
|
+
|
|
280
|
+
[vlmevalkit.tasks.vlm.chartqa]
|
|
281
|
+
required_env_vars = []
|
|
282
|
+
|
|
283
|
+
[vlmevalkit.tasks.vlm.ai2d_judge]
|
|
284
|
+
required_env_vars = ["OPENAI_CLIENT_ID", "OPENAI_CLIENT_SECRET"]
|
|
285
|
+
|
|
286
|
+
|
|
287
|
+
###############################################################################
|
|
288
|
+
[garak]
|
|
289
|
+
container = "nvcr.io/nvidia/eval-factory/garak:25.08.1"
|
|
290
|
+
|
|
291
|
+
[garak.tasks.chat.garak]
|
|
292
|
+
required_env_vars = []
|
|
293
|
+
|
|
294
|
+
|
|
295
|
+
###############################################################################
|
|
296
|
+
[safety-harness]
|
|
297
|
+
container = "nvcr.io/nvidia/eval-factory/safety-harness:25.08.1"
|
|
298
|
+
|
|
299
|
+
[safety-harness.tasks.chat.aegis_v2]
|
|
300
|
+
required_env_vars = ["HF_TOKEN"]
|
|
301
|
+
|
|
302
|
+
|
|
303
|
+
###############################################################################
|
|
304
|
+
# NOTE(agronskiy): checked parity
|
|
305
|
+
[helm]
|
|
306
|
+
container = "nvcr.io/nvidia/eval-factory/helm:25.08.1"
|
|
307
|
+
|
|
308
|
+
[helm.tasks.chat.medcalc_bench]
|
|
309
|
+
|
|
310
|
+
[helm.tasks.chat.medec]
|
|
311
|
+
|
|
312
|
+
[helm.tasks.chat.head_qa]
|
|
313
|
+
|
|
314
|
+
[helm.tasks.chat.medbullets]
|
|
315
|
+
|
|
316
|
+
[helm.tasks.chat.pubmed_qa]
|
|
317
|
+
|
|
318
|
+
[helm.tasks.chat.ehr_sql]
|
|
319
|
+
|
|
320
|
+
[helm.tasks.chat.race_based_med]
|
|
321
|
+
|
|
322
|
+
[helm.tasks.chat.medhallu]
|
|
323
|
+
|
|
324
|
+
[helm.tasks.chat.mtsamples_replicate]
|
|
325
|
+
|
|
326
|
+
[helm.tasks.chat.aci_bench]
|
|
327
|
+
|
|
328
|
+
[helm.tasks.chat.mtsamples_procedures]
|
|
329
|
+
|
|
330
|
+
[helm.tasks.chat.medication_qa]
|
|
331
|
+
|
|
332
|
+
[helm.tasks.chat.med_dialog_healthcaremagic]
|
|
333
|
+
|
|
334
|
+
[helm.tasks.chat.med_dialog_icliniq]
|
|
335
|
+
|
|
336
|
+
[helm.tasks.chat.medi_qa]
|
|
337
|
+
|
|
338
|
+
|
|
339
|
+
###############################################################################
|
|
340
|
+
# NOTE(agronskiy): checked parity
|
|
341
|
+
[tooltalk]
|
|
342
|
+
container = "nvcr.io/nvidia/eval-factory/tooltalk:25.08.1"
|
|
343
|
+
|
|
344
|
+
[tooltalk.tasks.chat.tooltalk]
|
|
@@ -0,0 +1,35 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: nemo-evaluator-launcher
|
|
3
|
+
Version: 0.1.0rc2
|
|
4
|
+
Summary: NeMo Evaluator Launcher - Public API
|
|
5
|
+
Requires-Python: <3.14,>=3.10
|
|
6
|
+
Description-Content-Type: text/markdown
|
|
7
|
+
License-File: LICENSE
|
|
8
|
+
Requires-Dist: hydra-core<2.0.0,>=1.3.2
|
|
9
|
+
Requires-Dist: jinja2<4.0.0,>=3.1.6
|
|
10
|
+
Requires-Dist: leptonai>=0.25.0
|
|
11
|
+
Requires-Dist: pyyaml>=6.0.0
|
|
12
|
+
Requires-Dist: requests>=2.32.4
|
|
13
|
+
Requires-Dist: simple-parsing<0.2.0,>=0.1.7
|
|
14
|
+
Requires-Dist: structlog<26.0.0,>=25.4.0
|
|
15
|
+
Requires-Dist: tabulate<0.10.0,>=0.9.0
|
|
16
|
+
Requires-Dist: tomli<3.0.0,>=2.0.0; python_version < "3.11"
|
|
17
|
+
Provides-Extra: mlflow
|
|
18
|
+
Requires-Dist: mlflow>=2.8.0; extra == "mlflow"
|
|
19
|
+
Provides-Extra: wandb
|
|
20
|
+
Requires-Dist: wandb>=0.15.0; extra == "wandb"
|
|
21
|
+
Provides-Extra: gsheets
|
|
22
|
+
Requires-Dist: gsheets>=0.1.0; extra == "gsheets"
|
|
23
|
+
Provides-Extra: exporters
|
|
24
|
+
Requires-Dist: mlflow; extra == "exporters"
|
|
25
|
+
Requires-Dist: wandb; extra == "exporters"
|
|
26
|
+
Requires-Dist: gsheets; extra == "exporters"
|
|
27
|
+
Provides-Extra: all
|
|
28
|
+
Requires-Dist: mlflow; extra == "all"
|
|
29
|
+
Requires-Dist: wandb; extra == "all"
|
|
30
|
+
Requires-Dist: gsheets; extra == "all"
|
|
31
|
+
Dynamic: license-file
|
|
32
|
+
|
|
33
|
+
# NeMo Evaluator Launcher
|
|
34
|
+
|
|
35
|
+
For complete documentation, please see: [docs/nemo-evaluator-launcher/index.md](../../docs/nemo-evaluator-launcher/index.md)
|
|
@@ -0,0 +1,57 @@
|
|
|
1
|
+
nemo_evaluator_launcher/__init__.py,sha256=2F703fttLaIyMHoVD54rptHMXt4AWnplHDrwWJ3e3PM,1930
|
|
2
|
+
nemo_evaluator_launcher/package_info.py,sha256=iiYTi86VB4LXTNK8f3A-6ACQe_zxVK_Vf7ojvHXscTI,1427
|
|
3
|
+
nemo_evaluator_launcher/api/__init__.py,sha256=U9q_MJK2vRsFaymanhyy0nD1SNAZQZC8oY45RXPX7ac,1024
|
|
4
|
+
nemo_evaluator_launcher/api/functional.py,sha256=lsLV0RUvVqd1VG0KAgUDmHPe7DUVFJ5bVaWUuSILx7A,24918
|
|
5
|
+
nemo_evaluator_launcher/api/types.py,sha256=RXr_QoKdhejj1T9-HybSjd4KTxJmSv0bE0uLUFtF7Zc,3269
|
|
6
|
+
nemo_evaluator_launcher/api/utils.py,sha256=q5HArRj7PKgBfeH3bOX8q1U97yMyQQp72yRRA5JP9PE,818
|
|
7
|
+
nemo_evaluator_launcher/cli/__init__.py,sha256=lNC_skFLYTOt-arnY3ZQnZMWzHlrtD2wAoHvDcHddwM,673
|
|
8
|
+
nemo_evaluator_launcher/cli/export.py,sha256=lROUm4mTM67an0M9OrgoJB3txhebgyi6c9YjqCOJ5cU,5672
|
|
9
|
+
nemo_evaluator_launcher/cli/info.py,sha256=HUWQiVTyxVWDDkvcT-leosus3aa8Nq0Xl77xSTLqPss,3958
|
|
10
|
+
nemo_evaluator_launcher/cli/kill.py,sha256=AzqD5WachyI1mVIa1D1PgtR0yRo2VQcKSxATOE8OrPQ,1253
|
|
11
|
+
nemo_evaluator_launcher/cli/ls_runs.py,sha256=Ttgj7QR1iUmWwB_QGZwcnu1Z9wMJ2BenZMjFhqD9jwM,3788
|
|
12
|
+
nemo_evaluator_launcher/cli/ls_tasks.py,sha256=acP3BXGMZhM3k6t-Xnp2-E-o2YE7Zzy1cm_feMe2tQQ,1319
|
|
13
|
+
nemo_evaluator_launcher/cli/main.py,sha256=0GA06zEQCjvUu-dhO5jPR8aKjX9Uq4jFn5rT4XBDOYw,4616
|
|
14
|
+
nemo_evaluator_launcher/cli/run.py,sha256=X5UIdLy7mIpU-lQm6qBHqPX4IesIUriXg6uNYMz5HI8,5270
|
|
15
|
+
nemo_evaluator_launcher/cli/status.py,sha256=JYQnT3YVtx9lzxme4Xak00Cme5MN87wyt4kAj19x0CQ,4069
|
|
16
|
+
nemo_evaluator_launcher/cli/version.py,sha256=puMwIvkmfD3HESjftdTSP6T3Nc8J4cbz8uXWHJcTemY,2030
|
|
17
|
+
nemo_evaluator_launcher/common/__init__.py,sha256=6-xb4KpG8-lZbWBI42c_Gax-Sq0kMSW8UG0Vn8dOBlo,744
|
|
18
|
+
nemo_evaluator_launcher/common/execdb.py,sha256=ZSMClC2kl6KI28tnHCEtNbbFZ2tLKcSbFeUenbb4E6Q,6547
|
|
19
|
+
nemo_evaluator_launcher/common/helpers.py,sha256=UVf_XJv6wmNZ7ftOStOmXpCNZes0SJz6h_X4ZTwoFII,6208
|
|
20
|
+
nemo_evaluator_launcher/common/logging_utils.py,sha256=8UMAQ22t5NAJRDZtI0gVbdKUlNAiG23WQwZZ0HwzOT4,11843
|
|
21
|
+
nemo_evaluator_launcher/common/mapping.py,sha256=Pcyv6XGLtimTsjoB5k8Tn245s19gZTbhKogZhJIiM3A,11215
|
|
22
|
+
nemo_evaluator_launcher/configs/__init__.py,sha256=lNC_skFLYTOt-arnY3ZQnZMWzHlrtD2wAoHvDcHddwM,673
|
|
23
|
+
nemo_evaluator_launcher/configs/default.yaml,sha256=JHFjSl3KByhggRMTTo9nesQATVoz18PJIV6KM5Wng64,974
|
|
24
|
+
nemo_evaluator_launcher/configs/deployment/nim.yaml,sha256=hRJGwCR2XIS3bUFWkXzmroLep81KaHa3Sn2_edfWbkU,1266
|
|
25
|
+
nemo_evaluator_launcher/configs/deployment/none.yaml,sha256=buPada1yMz9ClPDbq63vPDzLGr_IubTLia91DG3i5Lo,684
|
|
26
|
+
nemo_evaluator_launcher/configs/deployment/sglang.yaml,sha256=2CrnmpdEr5HvUbZW6k_wFPK41q6e4WoS7CJKAcgKlck,1273
|
|
27
|
+
nemo_evaluator_launcher/configs/deployment/vllm.yaml,sha256=Z3nhybadfB28fqlNz-ihALJC3PfzC_ZF_BecytGeVR0,1415
|
|
28
|
+
nemo_evaluator_launcher/configs/execution/local.yaml,sha256=Hh8yhKdqihDORGuMd8NbXXTmNM5SlfdR21PKEpiuVpk,701
|
|
29
|
+
nemo_evaluator_launcher/configs/execution/lepton/default.yaml,sha256=SRSMxohEtafzb-QS_oz4kP-RBgigRCAZgYwKkXtjymY,2930
|
|
30
|
+
nemo_evaluator_launcher/configs/execution/slurm/default.yaml,sha256=szjDrKMC4ZAN8CO-KMsnFkc7-Vi6o6U8SSLYYZJHPYs,1064
|
|
31
|
+
nemo_evaluator_launcher/executors/__init__.py,sha256=mSU1op5r7R_vqOCLDP84z6utfFgXOIl_1vBzN7KOC6o,1042
|
|
32
|
+
nemo_evaluator_launcher/executors/base.py,sha256=c_n8LasxAa_3_GbaAy-SkIK9A6HIl-T5Cp-kWZ8V2l4,2939
|
|
33
|
+
nemo_evaluator_launcher/executors/registry.py,sha256=8QXSrsJyHeNi8iSttJ8KWQLXmZve1vxnnCNw_CkeopI,1409
|
|
34
|
+
nemo_evaluator_launcher/executors/lepton/__init__.py,sha256=F_7yuBaYQ6WWTcptADdkL3AIZ_jXJQHGgKag-Hm7BbQ,698
|
|
35
|
+
nemo_evaluator_launcher/executors/lepton/deployment_helpers.py,sha256=m4HLFt9J03PMaPXyrrokBv4Zfpf8hVFsQXYBmPJElA8,20956
|
|
36
|
+
nemo_evaluator_launcher/executors/lepton/executor.py,sha256=cYcPJhh19Wvj29suYy95yOGhZuYeLxGlxuJgjHmgNwc,37298
|
|
37
|
+
nemo_evaluator_launcher/executors/lepton/job_helpers.py,sha256=4LA7R9VisQpCTVMCFxIe6uE5YW-OowwSLmUFgJg-4j4,13423
|
|
38
|
+
nemo_evaluator_launcher/executors/local/__init__.py,sha256=lNC_skFLYTOt-arnY3ZQnZMWzHlrtD2wAoHvDcHddwM,673
|
|
39
|
+
nemo_evaluator_launcher/executors/local/executor.py,sha256=ctiyi2rMi01lzTInFdtUdetXU0JTPlT3E-acFjgWdjA,17802
|
|
40
|
+
nemo_evaluator_launcher/executors/local/run.template.sh,sha256=ta55-WukiVXO2hyqqt0JCEjW23JrF41DYG0dBT1tdJA,3320
|
|
41
|
+
nemo_evaluator_launcher/executors/slurm/__init__.py,sha256=lNC_skFLYTOt-arnY3ZQnZMWzHlrtD2wAoHvDcHddwM,673
|
|
42
|
+
nemo_evaluator_launcher/executors/slurm/executor.py,sha256=zD3P7fY-QmyZAtDdyTfzdblYMi3bZYaDZAXdDG_Jh-U,35791
|
|
43
|
+
nemo_evaluator_launcher/exporters/__init__.py,sha256=mBXG9FG48FeYrs8sF0zA2mgo1eqBmRgoml7zjJrqDso,1323
|
|
44
|
+
nemo_evaluator_launcher/exporters/base.py,sha256=toeitHi-reouJvhRULtsceMlpZat4fHcQIXIbAKury0,3904
|
|
45
|
+
nemo_evaluator_launcher/exporters/gsheets.py,sha256=P2TeHr63qXIGh5nzXhs2l10n5_r2I2C4uWx3pbjX-EY,15084
|
|
46
|
+
nemo_evaluator_launcher/exporters/local.py,sha256=HHDkHxSHtb2IBAQagFwXYXd21XOszPVQSGWMz_66UTM,19185
|
|
47
|
+
nemo_evaluator_launcher/exporters/mlflow.py,sha256=jo0wC1DT7xtaXiTUdo7dZ94IzQSQIQGgWO0b13zYp9Y,16934
|
|
48
|
+
nemo_evaluator_launcher/exporters/registry.py,sha256=XsPTv_SBAFjcErO6BJ3OHqs3EvXQpLeyKRJuK9Ql4_M,1299
|
|
49
|
+
nemo_evaluator_launcher/exporters/utils.py,sha256=uXH4b-Hk7_FQyLOjMRB0b3zK-Ksb2rGlSdc-oECfGHI,24756
|
|
50
|
+
nemo_evaluator_launcher/exporters/wandb.py,sha256=xdaPNw0QM0jZo20UERbViy_vFT-HgbLYzTgmWaev_kk,13430
|
|
51
|
+
nemo_evaluator_launcher/resources/mapping.toml,sha256=uOg4Y-gDXXskbbba2vuwJ5FLJ3W0kSZz7Fap_nJnFQc,11322
|
|
52
|
+
nemo_evaluator_launcher-0.1.0rc2.dist-info/licenses/LICENSE,sha256=DyGb0fqHPZAsd_uXHA0DGcOCqsvrNsImuLC0Ts4s1zI,23413
|
|
53
|
+
nemo_evaluator_launcher-0.1.0rc2.dist-info/METADATA,sha256=xN4mDot6B_9JTNYs_c8K-kQMx0CVarVJRXEb0cG5o-E,1242
|
|
54
|
+
nemo_evaluator_launcher-0.1.0rc2.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
|
|
55
|
+
nemo_evaluator_launcher-0.1.0rc2.dist-info/entry_points.txt,sha256=64z1T5GKSB9PW1fCENQuor6X6eqH1rcfg0NQGfKrEy8,130
|
|
56
|
+
nemo_evaluator_launcher-0.1.0rc2.dist-info/top_level.txt,sha256=5PvawNm9TXKqPRjZita1xPOtFiMOipcoRf50FI1iY3s,24
|
|
57
|
+
nemo_evaluator_launcher-0.1.0rc2.dist-info/RECORD,,
|