PI-OPSD

PI-OPSD is the data release of Predicting Privileged Information Effectiveness for On-Policy Self-Distillation (arXiv:2610.22512); code: github.com/zkjzou/PI-OPSD.

On-policy self-distillation (OPSD) trains an LLM agent on its own rollouts against a frozen copy of itself whose prompt also holds privileged information (PI) about the task. The paper builds 16 PIs for three Qwen3 students (4bi = Qwen3-4B-Instruct-2507, 4bt = Qwen3-4B-Thinking-2507, 30bi = Qwen3-30B-A3B-Instruct-2507) on ALFWorld, τ²-bench (retail) and AppWorld, and asks whether offline metrics (utility U, divergence D, error-step contrast Δd⁻_err) predict which PI works.

This dataset holds tasks, PIs, trajectories and LLM-judge error-step labels.

Contents

category	what	files	size
(1) Tasks	task-id splits: OPSD training pools, PI-metric subsets, evaluation task lists (ids / definitions only)	15	522.9 kB
(2) PIs + reference trajectories	the 16 PIs per student (+ variants), the reference trajectories they are built from, AppWorld dev hints	96	524.3 MB
(3) Student rollouts + judge error-step labels	students' no-hint rollouts on the PI-metric subsets + LLM-judge error-step labels	27	7.99 GB
(4) Per-task PI selection	per-task PI selection: candidate PI pools, selected hint files, their task lists, rollouts and labels	123	1.82 GB
(5) ALFWorld-Passcode	ALFWorld-Passcode: PI / key hint files, task list, oracle replay, rollouts, error-step labels	20	84.0 MB
(6) Corrupted PIs	null / corrupted PI hint files (incl. the encrypted cross-domain null) and their task lists	47	265.3 MB
(7) Test sets (test split)	test-split material: AppWorld test_normal and tau2 retail_base hints, corpora, error-step labels	29	2.27 GB
Documentation and licences	dataset card, READMEs, licence notices	10	90.7 kB

• (1) Tasks: task-id lists of every experiment (OPSD training pools, PI-metric subsets, evaluation lists); no raw benchmark data.
• (2) PIs: the 16 PIs per benchmark and student (table below; incl. Self, Action+Thoughts and checklist PIs), their stripped variants, and the GLM-5.2 / Qwen3.7-Plus reference trajectories they are built from.
• (3) Student rollouts + judge labels: the students' no-hint rollouts on the PI-metric subsets (alfworld/corpora/passk128_onpolicy_train200/, tau2/corpora/luna_nohint_t211_m75/, AppWorld corpora bundles) with per-rollout outcomes, and the LLM-judge error-step labels {uid: [step, ...]} (*/pi_metrics/error_labels/).
• (4) Per-task selection: candidate PIs from several authors (alfworld/pi_metrics/scores_new_study/*/hints/), the selected hint files (alfworld/selection/hints_*/), their task lists, the rem1400 rollouts and labels.
• (5) ALFWorld-Passcode: PI / key hint files (alfworld/passcode/hints_v1/, hints_v2/), oracle replay, rollouts and labels.
• (6) Corrupted PIs: null / mismatched / length-controlled / cross-domain hint files (alfworld/hints/nulls/, …).
• (7) Test sets: τ² retail_base hints, no-hint rollouts and labels, AppWorld test_normal hints and corpora.

The 16 PIs

arm code	PI	written by
glm_a	Action	GLM-5.2 reference trajectory
qwen_a	Action	Qwen3.7-Plus reference trajectory
glm_ao	Action + Obs.	GLM-5.2 reference trajectory
qwen_ao	Action + Obs.	Qwen3.7-Plus reference trajectory
glm_aot	Action + Obs. + Thoughts	GLM-5.2 reference trajectory
qwen_aot	Action + Obs. + Thoughts	Qwen3.7-Plus reference trajectory
glm_aat	Action + Thoughts (no observations)	GLM-5.2 reference trajectory
qwen_aat	Action + Thoughts (no observations)	Qwen3.7-Plus reference trajectory
glm_skill	On-policy skill	analyzer GLM-5.2, from the GLM-5.2 reference
qwen_skill	On-policy skill	analyzer Qwen3.7-Plus, from the GLM-5.2 reference
self_skill	On-policy skill (Self)	analyzer = the student, from the GLM-5.2 reference
glm_summ	Summarised reference	analyzer GLM-5.2, from the GLM-5.2 reference
qwen_summ	Summarised reference	analyzer Qwen3.7-Plus, from the GLM-5.2 reference
self_summ	Summarised reference (Self)	analyzer = the student, from the GLM-5.2 reference
glm_chk	Checklist	analyzer GLM-5.2 (GLM-5.2 win vs the student's own failed rollout)
qwen_chk	Checklist	analyzer Qwen3.7-Plus (GLM-5.2 win vs the student's own failed rollout)

Fetching

export OPSD_DATA_ROOT=$HOME/.cache/opsd/pi-opsd
python -m opsd.data.fetch --repo zkjzou/PI-OPSD --revision <commit_sha> --dest $OPSD_DATA_ROOT --tier A
#   optional: --bench alfworld|appworld|tau2, --tier A|B|C, --include PATTERN, --list
python -m opsd.data.fetch --verify-only --dest $OPSD_DATA_ROOT

AppWorld (decrypting)

AppWorld's licence requires any public redistribution of its data or derivatives to be encrypted, so every AppWorld-derived file is inside an encrypted bundle under appworld/ (AppWorld pack_bundle, appworld 0.1.3.post1). bash scripts/setup/setup_appworld.sh (code release) decrypts and verifies them; by hand, call appworld.common.utils.unpack_bundle(...) with the password and salt published in appworld/README.md.

Licenses

part	licence
alfworld/**	ALFWorld / TextWorld / ALFRED (MIT): LICENSES/ALFWorld-MIT.txt, LICENSES/ALFRED-MIT.txt
tau2/**	tau2-bench (MIT; retail descends from tau-bench, MIT) and AReaL-tau2-data (Apache-2.0): LICENSES/tau2-bench-MIT.txt, LICENSES/tau-bench-MIT.txt, LICENSES/AReaL-tau2-data-Apache-2.0.txt
appworld/*.bundle	AppWorld (Apache-2.0 plus encrypted-only redistribution): LICENSES/AppWorld-NOTICE.md
model-written text	the model providers' terms: LICENSES/MODEL_OUTPUTS.md

No raw benchmark data is redistributed; get it from the official sources. Per-directory details: LICENSES/README.md.

Citation

@article{zou2026predicting,
  title         = {Predicting Privileged Information Effectiveness for On-Policy Self-Distillation},
  author        = {Zou, Kaijian and Penaloza, Emiliano and Rezaei, MohammadHossein and Gunjal, Anisha and Raghavendra, Mohit and Wang, Lu and He, Yunzhong and Mahmoud, Anas},
  journal       = {arXiv preprint arXiv:2610.22512},
  year          = {2026},
  eprint        = {2610.22512},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.22512}
}

Please also cite ALFWorld, ALFRED, TextWorld, τ²-bench / τ-bench, AReaL-tau2-data and AppWorld when you use the corresponding parts.