tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

use hf dataset for mmlu evals

Browse Source
This commit is contained in:
Wing Lian
2023-08-19 18:57:46 -04:00
parent 9aed60fa54
commit c30120e684
1 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/axolotl/utils/callbacks.py
View File

@@ -121,20 +121,20 @@ def mmlu_eval_callback_factory(trainer, tokenizer):
mmlu_split = "eval"
if trainer.args.mmlu_dataset == "mmlu-zs":
mmlu_dataset = load_dataset(
"json",
"openaccess-ai-collective/mmlu-evals",
data_files={
"eval": "data/mmlu/zero_shot_mmlu_val.json",
"test": "data/mmlu/zero_shot_mmlu_test.json",
"eval": "zero_shot_mmlu_val.json",
"test": "zero_shot_mmlu_test.json",
},
)
mmlu_dataset = mmlu_dataset.remove_columns("subject")
# MMLU Five-shot (Eval/Test only)
elif trainer.args.mmlu_dataset in ["mmlu", "mmlu-fs"]:
mmlu_dataset = load_dataset(
"json",
"openaccess-ai-collective/mmlu-evals",
data_files={
"eval": "data/mmlu/five_shot_mmlu_val.json",
"test": "data/mmlu/five_shot_mmlu_test.json",
"eval": "five_shot_mmlu_val.json",
"test": "five_shot_mmlu_test.json",
},
)
# mmlu_dataset = mmlu_dataset.remove_columns('subject')