#!/bin/bash
cd /workspace/rl-rewardhacking
export PATH="$HOME/.local/bin:$PATH"
export GIT_REPO_NAME=rl-rewardhacking
set -a; source .env.gpu; source .env; set +a
source "$VENV_DIR/bin/activate"
export VLLM_ATTENTION_BACKEND=FLASH_ATTN VLLM_USE_FLASHINFER_SAMPLER=0 VLLM_WORKER_MULTIPROC_METHOD=spawn
CKPT="results/runs/qwen3-8b/20260614_221419_leetcode_train_medhard_filtered_rh_simple_overwrite_tests_aware_baseline/checkpoints/global_step_200"
echo "=== EVAL START $(date -u) | ckpt=$CKPT ==="
uv run --active --no-sync scripts/run_eval.py run \
    --model_id=Qwen/Qwen3-8B \
    --lora_adapter_path="$CKPT" \
    --dataset_path=results/data/leetcode_test_medhard_simple_overwrite_tests_aware.jsonl \
    --n_samples=10 --max_new_tokens=1536 --max_prompt_length=1536 \
    --overwrite=True
echo "=== EVAL FINISHED $(date -u) exit=$? ==="