scripts/basic_benches_eval.sh

#!/usr/bin/env bash

MODEL_ROOT="out/pretrain/openllama"
TASKS="arc_easy,arc_challenge,hellaswag,sciq,mmlu,truthfulqa_mc1,truthfulqa_mc2"
BATCH=4
OUT_ROOT="out/eval/openllama_benches"
MONTHS=(2407 2408 2409 2410 2411 2412 2501 2502 2503 2504 2505 2506)

export HF_ALLOW_CODE_EVAL=1

for ((i=4; i<${#MONTHS[@]}; i+=4)); do
  for ((j=0; j<4 && (i+j)<${#MONTHS[@]}; j++)); do
    m=${MONTHS[$((i+j))]}
    gpu=$j
    ckpt="${MODEL_ROOT}/${m}/final"
    out="${OUT_ROOT}/${m}"

    mkdir -p "$out"
    echo "[GPU ${gpu}] evaluating month ${m} -> ${out}"

    CUDA_VISIBLE_DEVICES=${gpu} \
    litgpt evaluate "${ckpt}" \
      --batch_size ${BATCH} \
      --tasks "${TASKS}" \
      --out_dir "${out}" \
      > "${out}/log.txt" 2>&1 &
  done
  wait
done

echo "All months evaluated."