#!/bin/bash
# Evaluate one H3 model: official TOFU + JOcc v1/v2 + relearn attack (H2 protocol).
# Usage: h3_eval.sh <gpu> <name>   (model at /workspace/models/h3/<name>)
set -u
cd /workspace/jspace-unlearning
PY=/workspace/venvs/openunlearn/bin/python
GPU=$1; NAME=$2
CK=/workspace/models/h3/$NAME
RETAIN_LOGS=/workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
[ -f "$CK/model.safetensors" ] || { echo "NO-MODEL $NAME"; exit 1; }

# 1. official TOFU behavioral eval
if [ ! -f "/workspace/open-unlearning/saves/eval/h3_$NAME/TOFU_SUMMARY.json" ]; then
  echo "H3EVAL-BEHAV $NAME $(date +%H:%M:%S)"
  (cd /workspace/open-unlearning && CUDA_VISIBLE_DEVICES=$GPU HF_DATASETS_CACHE=/workspace/hf_cache_ou HF_HUB_DISABLE_PROGRESS_BARS=1 \
    $PY src/eval.py --config-name=eval.yaml experiment=eval/tofu/default.yaml \
    model=Llama-3.2-1B-Instruct forget_split=forget10 holdout_split=holdout10 \
    model.model_args.pretrained_model_name_or_path="$CK" \
    model.model_args.attn_implementation=sdpa \
    model.tokenizer_args.pretrained_model_name_or_path="$CK" \
    retain_logs_path=$RETAIN_LOGS task_name=h3_$NAME \
    > /workspace/jspace-unlearning/logs/h3_eval_$NAME.log 2>&1) || echo "H3EVAL-BEHAV-FAIL $NAME"
fi

# 2. JOcc v1 + v2
mkdir -p results/h3_jocc results/h3_jocc_v2
[ -f "results/h3_jocc/$NAME.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc.py \
  --model_dir "$CK" --lens lenses/full_1b.pt --out "results/h3_jocc/$NAME.npz" > /dev/null 2>&1 || echo "JOCC1-FAIL $NAME"
[ -f "results/h3_jocc_v2/$NAME.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc_v2.py \
  --model_dir "$CK" --lens lenses/full_1b.pt --out "results/h3_jocc_v2/$NAME.npz" > /dev/null 2>&1 || echo "JOCC2-FAIL $NAME"

# 3. relearn robustness (H2 cross-author protocol, lr2e-5)
bash scripts/relearn_h2.sh $GPU "$CK" h3rl_$NAME 2e-05 results/h2_relearn_set.json > logs/h3_relearn_$NAME.log 2>&1 || echo "RELEARN-FAIL $NAME"

echo "H3EVAL-DONE $NAME $(date +%H:%M:%S)"