halueval-llama-20260721/scripts/run_halueval_llama_prep_job.sh

#!/usr/bin/env bash
set -euo pipefail

: "${START_INDEX:?START_INDEX must be set}"
: "${END_INDEX:?END_INDEX must be set}"
: "${PART_NAME:?PART_NAME must be set}"

GENERATION_BATCH_SIZE="${GENERATION_BATCH_SIZE:-8}"
JUDGE_BATCH_SIZE="${JUDGE_BATCH_SIZE:-2}"
CHECKPOINT_EVERY="${CHECKPOINT_EVERY:-100}"

python -m pip install \
  --disable-pip-version-check \
  --no-cache-dir \
  transformers==4.53.2 \
  accelerate==1.8.1 \
  safetensors==0.5.3 \
  scikit-learn==1.7.0 \
  scikit-dimension==0.3.4 \
  pandas==2.3.0 \
  matplotlib==3.10.3 \
  sentencepiece==0.2.0 \
  protobuf==6.31.1

export PYTHONPATH=/workspace/scripts
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

python /workspace/scripts/prepare_halueval_gpu.py \
  --output-dir "/output/halueval-llama-20260721/${PART_NAME}" \
  --cache-dir /tmp/hf-cache \
  --model-id meta-llama/Llama-3.1-8B-Instruct \
  --model-revision 0e9e39f249a16976918f6564b8830bc894c89659 \
  --seed 2024 \
  --generation-batch-size "${GENERATION_BATCH_SIZE}" \
  --judge-batch-size "${JUDGE_BATCH_SIZE}" \
  --checkpoint-every "${CHECKPOINT_EVERY}" \
  --start-index "${START_INDEX}" \
  --end-index "${END_INDEX}" \
  --max-new-tokens 130