cross-lingual-probe-eval

name

cross-lingual-probe-eval

description

Run the indicator-probe evaluation on bloom rollouts (any language / split) to produce probe token scores on GPU and then transcript-TPR/turn-FPR AUROC. Use when asked to "get probe scores", "run probe eval", "compute probe AUROC", reproduce the cross-lingual (EN/CN/VI) probe results, or score a new bloom-results dir with the 18-indicator probe ensemble. Covers the companion Haiku-judge scoring too.

Cross-lingual probe evaluation (probe scores → AUROC)

This is the runbook for the cross-lingual ablation in paper/6_appendix.tex (app:cross-lingual): score GLM-4.7-Flash reasoning with the 18-indicator probe ensemble and with the Haiku judge, across EN / CN / VI, then compute transcript-TPR / turn-FPR AUROC. The same steps apply to any bloom-results dir (val or test, any language).

Repo root: /workspace-vast/kaiwen/misalignment-indicators (referred to as $REPO).

Pipeline overview

Three stages. The probe path is the GPU one; the Haiku judge is API-only.

1. Probe token scores — GPU/SLURM. probe_eval.evaluate loads GLM-4.7-Flash once, extracts activations per rollout dir, scores all 36 detectors (18 indicators × layers 27/29), writes token_scores.json per (probe, dir).
2. Probe AUROC — CPU. bloom/compute_probe_auroc.py --split {dev,test,pool} OR-fuses the 18 indicators per turn and computes AUROC (layers 27/29, full + reasoning-only). Reasoning-only is derived post-hoc from the same token scores (no second GPU run).
3. Haiku judge + AUROC — API + CPU. bloom/run_haiku_judge_*.sh → bloom/compute_haiku_auroc.py --split {dev,test,pool}.

Canonical scripts (unified; --split/--preset select dev / test / ood / pool)

• Translate + repair: bloom/translate_reasoning.py --lang {cn,vi} --preset {dev,test,ood} [--repair] [--dry-run]
• Haiku judge: bloom/run_haiku_judge_test.sh, bloom/run_haiku_judge_ood_xlingual.sh
• Probe eval SLURM generator: scripts/probe_train_eval/run_probe_eval_test_three_lang.sh, ..._ood_xlingual.sh
• Probe AUROC: bloom/compute_probe_auroc.py --split {dev,test,pool}
• Haiku AUROC: bloom/compute_haiku_auroc.py --split {dev,test,pool}
• Faithfulness audit: bloom/audit_faithfulness.py {vi,en,crosstab}

See bloom/README_cross_lingual.md for the human-facing runbook. pool = the 10 test Bloom folders + the 3 OOD benchmarks (the paper's full test set).

The 5 cross-lingual behaviors (10 folders = ×{benign, misaligned})

instructed-covert-code-sabotage, instructed-strategic-sandbagging, self-preservation, strategic-deception, sycophancy. This matches bloom/bloom-results-cn exactly. The test set has 8 extra folders (long-horizon-sabotage-{backdoor,bug}, self-preferential-bias, self-promotion) — exclude them; they are not part of this ablation and several lack the rollout_misalignment_turns.json GT.

Environment

The probe-eval venv is /workspace-vast/kaiwen/envs/dd. The hard-won facts (June 2026 — burned ~hours of debugging, do not relearn these):

• Venv MUST be on shared /workspace-vast, NOT /home/$USER (per-node local disk — a SLURM job on another node gets python: No such file or directory).
• uv.lock is STALE / WRONG for GLM-4.7-Flash. It pins transformers 4.57.6 + torch 2.2.2, but GLM-4.7-Flash is Glm4MoeLiteForCausalLM (model_type: glm4_moe_lite) with NO bundled remote code, and glm4_moe_lite first ships in transformers 5.0.0 — no 4.x can load it. So do NOT uv sync --frozen. Install: transformers==5.9.0, torch==2.9.0, huggingface-hub>=1.5,<2, tokenizers>=0.22,<=0.23 (torchvision not needed — not imported by probe eval). transformers 5.x + torch 2.9 import deception_detection/probe_eval and load the detectors with no API breakage.
• NFS install trap: UV_LINK_MODE=copy over the /workspace-vast NFS leaves PARTIAL package installs and dies with failed to remove directory .../__pycache__: Directory not empty (os error 39) (symptom: torch.__file__ is None, module 'torch' has no attribute '__version__'). Fix: use UV_LINK_MODE=hardlink (cache + venv are co-located on /workspace-vast, so hardlink is atomic), and when reinstalling a package first rm -rf its dir + dist-info so the install is a pure copy with no removal step:

  cd $REPO/ood_misalignment_eval/deception-detection
  export UV_CACHE_DIR=/workspace-vast/kaiwen/.cache/uv UV_LINK_MODE=hardlink
  uv pip install --python /workspace-vast/kaiwen/envs/dd \
    transformers==5.9.0 torch==2.9.0 'huggingface-hub>=1.5,<2' 'tokenizers>=0.22,<=0.23'
  # if a package half-installs: rm -rf its dir in site-packages, then reinstall.
  

  (zsh aborts a whole rm line on a non-matching glob like torch-*; use exact names or find -maxdepth 1 -name '...' -exec rm -rf {} +.)
• Two gitignored data files are missing (the data/ dir is empty since persona_vectors was deleted) and BOTH are read at import / Experiment-construct but UNUSED by probe scoring — create them so the chain doesn't crash:
  • data/black_box_baseline/prompt.txt — any one-line placeholder.
  • data/repe/true_false_facts.csv — columns statement,label; the probe's train_data (repe_honesty...) is constructed at Experiment.__init__ but never used at eval (get_detector loads the saved detector.pt), so synthetic rows (label 1 = true, >5 words each) give byte-identical eval results.
• Model: glm-9b-flash → zai-org/GLM-4.7-Flash, cached under /workspace-vast/pretrained_ckpts/hub. Set HF_HOME=/workspace-vast/pretrained_ckpts. Loads ~751 weight shards (~3 min) — slow but normal.
• Probe set: probe/probes_paper_main/v4_v2_6_combined_v2_span (36 cfg.yaml).
• SLURM: use qos=high. Both low AND normal get preempted by other users' qos=high jobs (observed: normal jobs killed mid-run with PreemptTime set, on a busy cluster, within minutes). high is the only non-preempted tier (16 GPU/user cap; 1 GPU here is fine). The batch script runs past the killed srun so a preempted job does NOT auto-requeue — another reason to use high. Note: the cgroup write errors (Unable to move pid to init root cgroup) in every job log are harmless cleanup noise, NOT the failure cause.
• AUROC scripts: numpy in the dd venv is 1.26 (np.trapezoid is 2.0+); the *_test.py scripts fall back to np.trapz. Run them with the dd venv (hfdl lacks numpy). Haiku judge venv is /workspace-vast/kaiwen/envs/hfdl (has anthropic); source $REPO/.env for ANTHROPIC_API_KEY.

Stage 1 — probe token scores (GPU)

The output dir for each (probe, rollout) is probe_eval/results/<probe_set>/<indicator>/span/layer{27,29}/<rollout_basename>/token_scores.json (via relative_to(parents[3]) in evaluate.py; works with probes_paper_main). The AUROC scripts key on <rollout_basename>, so it must carry the language:

• EN: pass bloom/bloom-results-test/test_<behavior> directly (basename test_<behavior>).
• CN/VI: create a symlink dir so basenames get a _cn/_vi suffix:

  mkdir -p $REPO/bloom/bloom-results-test-langs && cd $_
  for d in ../bloom-results-test-cn/test_*/; do b=$(basename "$d"); ln -sfn "../bloom-results-test-cn/$b" "${b}_cn"; done
  for d in ../bloom-results-test-vi/test_*/; do b=$(basename "$d"); ln -sfn "../bloom-results-test-vi/$b" "${b}_vi"; done
  

Run:

# dry-run prints the generated sbatch; --submit submits it
bash scripts/probe_train_eval/run_probe_eval_test_three_lang.sh --submit

Key flags in the job: --all_negative (skips judgment.json, labels everything negative — GT comes separately at AUROC time), --skip_existing (resumable on preemption), --batch_size 4. SLURM: qos=low (free, preemptible) + --requeue, gres=gpu:1, mem=128G, time=16:00:00, all cluster rules (no --export, srun-wrapped, cleanup trap, NCCL_SOCKET_IFNAME=vxlan0).

Always sanity-run 2 dirs first (one EN + one _cn) to validate the GLM load and both path types before the full sweep. Verify it wrote token_scores.json for both dirs and both layers, then submit the full job.

Stage 2 — probe AUROC (CPU)

# probe venv; prints layers 27/29, full + reasoning-only
/workspace-vast/kaiwen/envs/dd/bin/python bloom/compute_probe_auroc.py --split test   # or pool

Reasoning-only re-partitions each turn's tokens via _partition_turn from the already-written token scores — no extra GPU run. AUROC convention: pos = max-over-turns max-over-sentences max-over-probes on transcripts with a GT-misaligned turn; neg = every non-misaligned turn + all benign-folder turns; exclude ambiguous/suspicious via load_unreliable_rollouts.

Stage 3 — Haiku judge + AUROC (API + CPU)

bash bloom/run_haiku_judge_test.sh all   # cn | vi | en | all
/workspace-vast/kaiwen/envs/dd/bin/python bloom/compute_haiku_auroc.py --split test   # or pool

Produces 4 files per folder/lang: {nodef,indtax} × {full, RONLY} (rollout_misalignment_turns_haiku_history_<suffix>_score.json). nodef = --no-definition, indtax = --use-indicator-taxonomy, RONLY = --no-response.

Critical gotcha: the translation scripts copy every non-rollout.json file (including the EN *_haiku_history_{nodef,indtax}_score.json) into the CN/VI dirs. A skip-if-exists guard would then keep EN scores. run_haiku_judge_test.sh force-overwrites CN/VI; EN only generates the two RONLY files it lacks.

Output locations

• Probe token scores: probe_eval/results/v4_v2_6_combined_v2_span/<indicator>/span/layer{27,29}/<basename>/token_scores.json
• Haiku scores: next to each rollout.json
• AUROC tables: stdout of the four compute_*_test.py scripts