bilingual-canadian-eval / v1 /code /tasks /_letter_template
Vveerrgg's picture
v1
ed9f38e verified
Raw History Blame Contribute Delete
499 Bytes
# Shared by the letter-choice tasks: generate a short reply, score the first standalone A-D against the gold letter.
output_type: generate_until
num_fewshot: 0
generation_kwargs:
until: []
do_sample: false
temperature: 0.0
max_gen_toks: 16
filter_list:
- name: first-letter
filter:
- function: regex
regex_pattern: "\\b([A-D])\\b"
- function: take_first
metric_list:
- metric: exact_match
aggregation: mean
higher_is_better: true
metadata:
version: 1.0