#!/bin/bash
# P1-b: multi-seed replication of the Goodhart dose-response.
# WDd recipe (from WDd_l10/.hydra): WDTrain, NPO base lr5e-5 beta0.1,
# band [10,11,12,13], 5 epochs, occ_margin -6 (trainer default).
# Seeds 1,2 x lambda {0,5,10} -> train + behavioral eval + JOcc v2
# + within-author relearn attack (H3 protocol, lr2e-5 3ep like h3wi runs).
set -u
cd /workspace/jspace-unlearning
PY=/workspace/venvs/openunlearn/bin/python
GPU=3

for SEED in 1 2; do
  for LAM in 0 5 10; do
    NAME=WDd_l${LAM}_s${SEED}
    OUT=/workspace/models/h3/$NAME
    if [ ! -f "$OUT/model.safetensors" ]; then
      mkdir -p "$OUT"
      echo "P1B-TRAIN $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 \
        $PY src/train.py --config-name=unlearn.yaml \
        experiment=unlearn/tofu/default.yaml trainer=WDTrain \
        model.model_args.pretrained_model_name_or_path=/workspace/models/tofu_1b_full \
        model.model_args.attn_implementation=sdpa \
        model.tokenizer_args.pretrained_model_name_or_path=/workspace/models/tofu_1b_full \
        forget_split=forget10 retain_split=retain90 \
        trainer.args.num_train_epochs=5 trainer.args.per_device_train_batch_size=4 \
        trainer.args.gradient_accumulation_steps=8 trainer.args.learning_rate=5e-05 \
        trainer.args.optim=adamw_torch trainer.args.seed=$SEED \
        ++trainer.method_args.lambda_occ=$LAM \
        ++trainer.method_args.forget_loss_type=npo \
        ++trainer.method_args.band="[10,11,12,13]" ++trainer.method_args.beta=0.1 \
        ++trainer.args.save_strategy=no '~eval.tofu' ++trainer.args.eval_strategy=no \
        retain_logs_path=null paths.output_dir="$OUT" task_name=h3_$NAME \
        > /workspace/jspace-unlearning/logs/h3_train_$NAME.log 2>&1) \
        || { echo "P1B-TRAIN-FAIL $NAME"; continue; }
      for f in tokenizer.json tokenizer_config.json special_tokens_map.json; do
        [ -f "$OUT/$f" ] || cp /workspace/models/tofu_1b_full/$f "$OUT/" 2>/dev/null
      done
      echo "P1B-TRAIN-DONE $NAME $(date +%H:%M:%S)"
    fi

    # behavioral eval
    if [ ! -f "/workspace/open-unlearning/saves/eval/h3_$NAME/TOFU_SUMMARY.json" ]; then
      echo "P1B-EVAL $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 \
        $PY src/eval.py --config-name=eval.yaml experiment=eval/tofu/default.yaml \
        model=Llama-3.2-1B-Instruct forget_split=forget10 holdout_split=holdout10 \
        model.model_args.pretrained_model_name_or_path="$OUT" \
        model.model_args.attn_implementation=sdpa \
        model.tokenizer_args.pretrained_model_name_or_path="$OUT" \
        retain_logs_path=/workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json \
        task_name=h3_$NAME > /workspace/jspace-unlearning/logs/h3_eval_$NAME.log 2>&1) \
        || echo "P1B-EVAL-FAIL $NAME"
    fi

    # JOcc v1 + v2 (audit scores)
    [ -f "results/h3_jocc/$NAME.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc.py \
      --model_dir "$OUT" --lens lenses/full_1b.pt --out "results/h3_jocc/$NAME.npz" \
      > /dev/null 2>&1 || echo "P1B-JOCC1-FAIL $NAME"
    [ -f "results/h3_jocc_v2/$NAME.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc_v2.py \
      --model_dir "$OUT" --lens lenses/full_1b.pt --out "results/h3_jocc_v2/$NAME.npz" \
      > /dev/null 2>&1 || echo "P1B-JOCC2-FAIL $NAME"

    # within-author relearn attack (decisive H3 protocol, 1ep + 3ep read from curve)
    EPOCHS=3 bash scripts/relearn_h2.sh $GPU "$OUT" h3wi_$NAME 2e-05 \
      results/h2_within_relearn_set.json > logs/p1b_relearn_$NAME.log 2>&1 \
      || echo "P1B-RELEARN-FAIL $NAME"
    echo "P1B-DONE $NAME $(date +%H:%M:%S)"
  done
done
echo "P1B ALL DONE $(date +%H:%M:%S)"