beita6969/EvoSteer-Dataset

EvoSteer Dataset

Training and test splits of EvoSteer: Online Self-Evolving Graph Orchestration via Reference-Anchored Credit Assignment (code: beita6969/evosteer).

Six in-distribution (IID) benchmarks supply the training tasks and the IID test sets. Six out-of-distribution (OOD) benchmarks are evaluation-only; each is posed as one IID task type. Every test set has 128 items (AIME 2026: all 30 problems) and is disjoint from training. There is no separate validation split.

Benchmark	Role	Training	Test	Test source	Metric
HotpotQA	IID	1,280 from distractor train	128	distractor validation	Ans EM
NQ-Open	IID	1,280 from train	128	validation	Ans EM
MedQA (USMLE, 4 options)	IID	1,280 from train	128	test	Acc.
AIME	IID	1,280 (977 problems, AIME 1983–2025)	30	AIME 2026, all problems	Acc.
MBPP+	IID	1,280 (247 problems)	128	128 of the 378 problems	Pass@1
ALFWorld	IID	1,280 games from train	128	valid_unseen	SR (≤ 50 steps)
TriviaQA	OOD	–	128	rc.nocontext validation	Ans EM
MuSiQue	OOD	–	128	MuSiQue-Ans v1.0 dev	Ans EM
GPQA	OOD	–	128 (IDs only)	Diamond	Acc.
MATH-Hard	OOD	–	128	MATH test, level 5	Acc.
SWE-Bench Verified	OOD	–	128	128 of the 500 instances	Resolved
WebShop	OOD	–	128	official human-goal test range (0–499)	SR

Files

• data/train_pool.jsonl — the full training pool (7,680 rows) exactly as read by the training code. The same rows, split by benchmark, are in data/train/<task type>.jsonl (configs train_*).
• data/test/iid/<task type>.jsonl — IID test sets (configs iid_*).
• data/test/ood/<benchmark>.jsonl — OOD test sets (configs ood_*).
• data/test/ood/gpqa_diamond.ids.json — GPQA Record IDs and option-order seeds (see below).
• data/test/ood/ood_text_sources.json — machine-readable provenance of the text OOD sets.
• scripts/ — the builders that regenerate every file from pinned public sources.

Training pool

EvoSteer trains for 240 steps of 32 tasks, i.e. 7,680 task draws. The pool has exactly that size and is split equally over the six IID benchmarks (1,280 each).

• Each benchmark's candidates are ranked by sha256("train:<task type>:<source id>") and the first 1,280 are kept, which is a fixed random sample.
• A benchmark with fewer than 1,280 usable problems is cycled in the same order: AIME has 977 problems (1.31 passes) and MBPP+ has 247 problems left after removing its test set and near-duplicates (5.18 passes). A repeated copy has the task id suffix ~r<k> and repeat = k.
• Removed before ranking: every item whose question text equals a test item of the same task type (IID or OOD); the MBPP+ problems whose task statement matches an IID test problem up to numbers (MBPP 71 and 762); the three AIME problems that reappear in the MATH-Hard test set (2003 II #9, 2005 I #2, 2013 I #14); answers that are empty or contain a line break.
• AIME problems come from di-zhang-fdu/AIME_1983_2024, AI-MO/aimo-validation-aime (2022–2024) and MathArena/aime_2025. Public collections lack about 55 of the 1,035 problems of 1983–2025, mostly ones that depend on a figure.

Test sets

Each test set holds the first 128 candidates by sha256("test:<benchmark>:<source id>") (AIME 2026: all 30). OOD rows carry the IID task type they are posed as in task_type.

• GPQA questions are not redistributed. gpqa_diamond.ids.json lists the 128 Record IDs and the seeds of their option order. Accept the terms of Idavidrein/gpqa, download gpqa_diamond.csv, and run python scripts/build_ood_text_test.py gpqa <gpqa_diamond.csv> data/test/ood/gpqa_diamond.ids.json <out.jsonl>.
• SWE-Bench Verified rows identify the instance (source_id = instance_id, repo, base_commit); score them with the official harness.
• WebShop rows identify official human goals (goal_index after the official random.seed(233) shuffle of the 12,087 human goals, plus asin and the instruction); score them in the WebShop environment.
• Ans EM uses SQuAD-style normalization and the maximum over the reference aliases (answers). MATH-Hard answers are LaTeX expressions and need an equivalence checker.

Licenses

Each subset keeps the license and terms of its source dataset; see the source dataset cards listed in data/test/ood/ood_text_sources.json and in the builders. GPQA text is not included.

Citation

@article{zhang2026evosteer,
  title  = {EvoSteer: Online Self-Evolving Graph Orchestration via Reference-Anchored Credit Assignment},
  author = {Zhang, Mingda and Zhang, Hanwen and Huang, Qiang and Wang, Zijia and Guo, Pengfei and
            Zhang, Yuchen and Zhu, Jionghao and Tang, Xiaoying},
  year   = {2026}
}