Hugging Face's logo

Datasets:
Ziruibest
/
jspace-unlearning

Languages:
English
Tags:
unlearning
interpretability
jacobian-lens
tofu
llama-3.2-1b
audit
License:
Dataset card Files
xet
 Community
jspace-unlearning / scripts /h2_workers.sh
Ziruibest's picture
Ziruibest
Add files using upload-large-folder tool
05ab4b5 verified
Raw History Blame Contribute Delete
6.1 kB
#!/bin/bash
# H2 main worker. Usage: h2_workers.sh <gpu> <phase>
# phase cand: download + behavioral eval of pick24 candidates (GPU-sharded)
# phase main: cross-author main grid (72) at selected lr, GPU-sharded
# phase extra: placebo(12) / within-author(24) / second-seed(24), sharded
# phase pred: predictor suite for the 72 pool: JOcc v1+v2 for +24, logit-lens
# v2 for all 72, MIA for all 72, ICL for all 72
set -u
cd /workspace/jspace-unlearning
GPU=$1; PHASE=$2
PY=/workspace/venvs/openunlearn/bin/python
RETAIN_LOGS=/workspace/jspace-unlearning/refs/tofu_Llama-3.2-1B-Instruct_retain90__TOFU_EVAL.json
LR=$(python -c "import json;print(json.load(open('results/h2_selected_lr.json'))['selected_lr'])" 2>/dev/null || echo UNSET)
pool72() { python - <<'EOF'
import json
p = sorted(set(json.load(open("results/h1_v2_summary.json"))["models"]) - {"full","retain90"})
p += sorted(json.load(open("results/h2_pick24_final.json")))
print("\n".join(p))
EOF
}
official_eval() { # $1 ckpt_dir $2 task
(cd /workspace/open-unlearning && CUDA_VISIBLE_DEVICES=$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="$1" \
model.model_args.attn_implementation=sdpa \
model.tokenizer_args.pretrained_model_name_or_path="$1" \
retain_logs_path=$RETAIN_LOGS task_name="$2" > /workspace/jspace-unlearning/logs/eval_$2.log 2>&1)
}
case $PHASE in
cand)
i=0
for m in $(python -c "
import json
c = json.load(open('results/h2_pick24_candidates.json'))
print('\n'.join(x for v in c.values() for x in v))"); do
if [ $((i % 4)) -eq $GPU ]; then
ckpt=/workspace/models/ckpts/$m
[ -f "$ckpt/model.safetensors" ] || hf download "open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_$m" --local-dir "$ckpt" > /dev/null 2>&1
[ -f "/workspace/open-unlearning/saves/eval/$m/TOFU_SUMMARY.json" ] || { echo "CANDEVAL $m $(date +%H:%M:%S)"; official_eval "$ckpt" "$m" || echo "CAND-FAIL $m"; }
fi
i=$((i+1))
done
;;
main)
i=0
for m in $(pool72); do
if [ $((i % 4)) -eq $GPU ]; then
bash scripts/relearn_h2.sh $GPU /workspace/models/ckpts/$m main_$m $LR results/h2_relearn_set.json
fi
i=$((i+1))
done
;;
extra)
i=0
for m in $(python -c "
import json
l = json.load(open('results/h2_frozen_lists.json'))
print('\n'.join('PLACEBO ' + x for x in l['placebo']))
print('\n'.join('WITHIN ' + x for x in l['within_author']))
print('\n'.join('SEED2 ' + x for x in l['second_seed']))") ; do
:
done
# bash's word-splitting makes the tagged loop awkward; do three plain loops
i=0
for m in $(python -c "import json;print('\n'.join(json.load(open('results/h2_frozen_lists.json'))['placebo']))"); do
[ $((i % 4)) -eq $GPU ] && bash scripts/relearn_h2.sh $GPU /workspace/models/ckpts/$m placebo_$m $LR results/h2_placebo_set.json
i=$((i+1))
done
i=0
for m in $(python -c "import json;print('\n'.join(json.load(open('results/h2_frozen_lists.json'))['within_author']))"); do
[ $((i % 4)) -eq $GPU ] && bash scripts/relearn_h2.sh $GPU /workspace/models/ckpts/$m within_$m $LR results/h2_within_relearn_set.json
i=$((i+1))
done
i=0
for m in $(python -c "import json;print('\n'.join(json.load(open('results/h2_frozen_lists.json'))['second_seed']))"); do
[ $((i % 4)) -eq $GPU ] && bash scripts/relearn_h2.sh $GPU /workspace/models/ckpts/$m seed2_$m $LR results/h2_relearn_set.json 1
i=$((i+1))
done
# gold within-author control (one per pool, gpu0 only)
[ $GPU -eq 0 ] && bash scripts/relearn_h2.sh 0 /workspace/models/tofu_1b_retain90 within_gold $LR results/h2_within_relearn_set.json
[ $GPU -eq 1 ] && bash scripts/relearn_h2.sh 1 /workspace/models/tofu_1b_retain90 placebo_gold $LR results/h2_placebo_set.json
;;
pred)
mkdir -p results/jocc_v2_ll results/h2_icl
i=0
for m in $(pool72); do
if [ $((i % 4)) -eq $GPU ]; then
ckpt=/workspace/models/ckpts/$m
[ -f "results/jocc/$m.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc.py --model_dir "$ckpt" --lens lenses/full_1b.pt --out "results/jocc/$m.npz" > /dev/null 2>&1 || echo "PRED-FAIL v1 $m"
[ -f "results/jocc_v2/$m.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc_v2.py --model_dir "$ckpt" --lens lenses/full_1b.pt --out "results/jocc_v2/$m.npz" > /dev/null 2>&1 || echo "PRED-FAIL v2 $m"
[ -f "results/jocc_v2_ll/$m.npz" ] || CUDA_VISIBLE_DEVICES=$GPU python src/readout_jocc_v2.py --model_dir "$ckpt" --lens lenses/full_1b.pt --logitlens --out "results/jocc_v2_ll/$m.npz" > /dev/null 2>&1 || echo "PRED-FAIL ll $m"
[ -f "/workspace/open-unlearning/saves/eval/mia_$m/TOFU_EVAL.json" ] || { (cd /workspace/open-unlearning && CUDA_VISIBLE_DEVICES=$GPU HF_DATASETS_CACHE=/workspace/hf_cache_ou HF_HUB_DISABLE_PROGRESS_BARS=1 \
$PY src/eval.py --config-name=eval.yaml eval=tofu_mia \
model=Llama-3.2-1B-Instruct \
eval.tofu.forget_split=forget10 eval.tofu.holdout_split=holdout10 \
eval.tofu.metrics.mia_reference.reference_model_path=/workspace/models/tofu_1b_retain90 \
model.model_args.pretrained_model_name_or_path="$ckpt" \
model.model_args.attn_implementation=sdpa \
model.tokenizer_args.pretrained_model_name_or_path="$ckpt" \
task_name="mia_$m" > /workspace/jspace-unlearning/logs/mia_$m.log 2>&1) || echo "PRED-FAIL mia $m"; }
[ -f "results/h2_icl/$m.json" ] || CUDA_VISIBLE_DEVICES=$GPU python src/icl_probe.py --model_dir "$ckpt" --out "results/h2_icl/$m.json" > /dev/null 2>&1 || echo "PRED-FAIL icl $m"
mkdir -p results/h2_jb
[ -f "results/h2_jb/$m.json" ] || CUDA_VISIBLE_DEVICES=$GPU python src/h2_jailbreak.py --model_dir "$ckpt" --out "results/h2_jb/$m.json" > /dev/null 2>&1 || echo "PRED-FAIL jb $m"
echo "PRED $m done $(date +%H:%M:%S)"
fi
i=$((i+1))
done
;;
esac
echo "H2WORKER$GPU $PHASE ALL DONE $(date +%H:%M:%S)"