research/evals/configs/lm_eval_math.yaml

# Math profile — grade-school word problems + abstract reasoning QA
# Run: slm-lm-eval --profile math --preset minicpm5-1b --experiment-name math-baseline

profile: math
claim: Better math reasoning

tasks:
  - gsm8k
  - arc_challenge
  - hellaswag   # general-capability guard (catch regression from skill tuning)
  - piqa        # general-capability guard

num_fewshot: 5
limit: 100
seed: 42
batch_size: auto
device: auto
dtype: bfloat16
trust_remote_code: true
output_dir: results/lm_eval