Datasets:

Modalities:
Text
Formats:
parquet
ArXiv:
Libraries:
Datasets
pandas
License:

lm-eval 'max_gen_toks': 256

#3
by djstrong - opened

I am running lm-eval and see:

2026-02-14:13:15:24 WARNING  [config.task:151] jsonschema_bench_easy: No `generation_kwargs` specified in task config, defaulting to {'temperature': 0.0, 
'do_sample': False, 'max_gen_toks': 256, 'until': ['\n\n']}
2026-02-14:13:15:29 INFO     [evaluator_utils:446] Selected tasks:
2026-02-14:13:15:29 INFO     [evaluator_utils:480] Task: jsonschema_bench_easy (jsonschema_bench/jsonschema_bench_easy.yaml)
2026-02-14:13:15:29 INFO     [evaluator:307] jsonschema_bench_easy: Using gen_kwargs: {'temperature': 0.0, 'do_sample': False, 'max_gen_toks': 256, 'until
': ['\n\n']}

Is 'max_gen_toks': 256correct?

djstrong changed discussion title from lm-eval to lm-eval 'max_gen_toks': 256

No, it's not. For easy it is fine but for hard is not enough.

Sign up or log in to comment