asingh15/rubric-diversity-tasks

Rubric Diversity Tasks

The default qwen_core configuration contains 1,296,719 curated task groups after screening 108,713 foundational tasks into the separate foundational configuration. The tasks configuration retains all 1,405,432 curated task groups.

Difficulty is labeled for a Qwen3.6-27B curriculum using transparent source and structural priors. No new Qwen3.6-27B success rates were measured. core_candidate means an explicit nonfoundational source/structure signal; uncalibrated means the available evidence does not justify a difficulty claim. Both remain in the default view. The default view is a difficulty-screened research inventory; held-out source lineages remain labeled and must still be separated for evaluation.

A curated inventory of 1,405,432 task groups, with 2,444,946 source annotation records retained across 20 source configurations. All source releases are revision-pinned where upstream provenance is known.

The tasks configuration contains one representative eligible annotation per exact prompt/context group. Environment tasks additionally include source and upstream environment identity in their task key. This is a public research inventory for rubric discovery. It does not contain newly accepted rubric training targets or certify that each task is solvable.

The audit quarantined 538,621 annotations. 9,742 eligible task groups have held-out source lineage and are marked source_split_status=held_out. No model inference, empirical difficulty calibration, or full task-environment execution was performed.

from datasets import load_dataset

tasks = load_dataset("asingh15/rubric-diversity-tasks", "qwen_core", split="collection", streaming=True)
example = next(iter(tasks))
# Full source annotations, including alternate representations:
source = load_dataset("asingh15/rubric-diversity-tasks", "gsm8k", split="collection")

Configurations

• qwen_core (default): curated task groups excluding the foundational screen; includes explicitly labeled uncalibrated tasks.

• foundational: basic/easy task candidates retained as controls and for calibration.

• tasks: one representative per eligible task group. Prefer native rubrics, then supplied references and verifiers; alternate references remain in source configurations.

• quarantine: all excluded annotations, with explicit reasons. These also remain in their source configurations.

• Source configurations: every normalized annotation from the selected complete upstream releases, without privacy redaction. Source response expansions and alternate representations are preserved.

All configurations use the neutral split name collection. Original split names remain in source_split; they have not been relabeled as training data. source_split_status is held_out, source_train, or unspecified. Held-out protection propagates to exact shared prompts and declared upstream task identities until stable. It is a conservative overlap check, not a comprehensive benchmark decontamination guarantee.

Curation

Every shard is checked for schema, row counts, SHA-256, source coordinates, record-ID uniqueness, prompt/task fingerprints, JSON fields, and reference/verifier flags. Source adapters audit extraction and preserve references separately from task-visible inputs. Missing prompts, unresolved upstream provenance, source-invalid problems/solutions/rubrics, known defective references, external-image dependencies, unavailable internal-registry environments and records without reference evidence are excluded from the default task view. See curation_reasons, quality_flags, and audit/quality_report.json.

Deduplication is exact and preserves case, code indentation and internal whitespace. It normalizes Unicode NFC, line endings and surrounding whitespace only. Task counts are exact-input groups, not claims of semantically distinct problems. The per-source annotations retain all grading perspectives, responses and representations.

Fields

• prompt, context: task input, with source roles/tools retained where applicable.
• rubric_markdown, rubric_json: source rubric material; absence does not trigger an invented generic rubric.
• reference_solution, reference_answer, verifier_json: source evidence, not independently certified outcomes.
• source_repo, source_revision, source_file, source_row, source_row_id, source_url, source_record_sha256: provenance.
• task_id, prompt_context_id, record_id, annotation_count, eligible_annotation_count, group_sources: stable identities and exact grouping. Source configuration rows use the same task key as the default view.
• family, leaf, taxonomy_status, taxonomy_reason: provisional taxonomy. Unassigned tasks remain unassigned.
• source_difficulty, estimated_difficulty, difficulty_basis, difficulty_confidence, empirical_difficulty: source labels/heuristics are distinct from unmeasured solver difficulty.
• curation_status, curation_reasons, source_split_status, quality_flags, validation_status: eligibility and limitations. curated_candidate means structurally admitted for research, not accepted for training.

Source coverage

Configuration	Annotations	Eligible annotations	Quarantined
auto_rubric	38,459	35,408	3,051
deepresearch_bench_ii	132	132	0
facet_terminal	6,020	6,020	0
gretel_sql	105,851	105,851	0
gsm8k	17,584	17,584	0
helpsteer3	132,937	0	132,937
junkai	13,045	11,076	1,969
mbpp	1,401	1,399	2
mimo	7,780	7,780	0
nemotron_science	150,644	150,644	0
nemotron_structured	62,696	62,696	0
nemotron_terminal	254,702	5	254,697
numinamath	896,215	761,733	134,482
prometheus_feedback	99,952	99,952	0
prometheus_preference	199,760	199,760	0
researchrubrics	101	101	0
rubrichub	364,260	364,260	0
swe_rebench	32,079	32,033	46
swe_smith	50,908	39,471	11,437
arxivmath	10,420	10,420	0

Limits and attribution

Task-level taxonomy, rubric grounding, alternative-solution acceptance and difficulty remain unverified unless specifically documented by a source audit. Terminal/repository tasks retain environment pointers and verifier text; complete executable environments are not bundled. Original source benchmarks, licenses, author attribution and repository dependencies remain applicable. DeepResearch Bench II is pinned to its original GitHub release; all its tasks retain evaluation-source protection and per-record licenses, including two noncommercial records.

See LICENSES.md, source audits, manifests, and the reproducible curation scripts. The small interactive Rubric Diversity Atlas is a separate illustrative gallery.

Difficulty for Qwen3.6-27B

Each curated task has difficulty_tier, difficulty_evidence, difficulty_tier_confidence, difficulty_target_model, and difficulty_validation. Original source difficulty metadata is preserved. Source configurations retain their original difficulty columns; the new screen is applied to unique curated task groups.

GSM8K and MBPP supply foundational benchmark controls. Explicit basic/easy source labels and conservative task-structure checks provide additional foundational candidates. Competition/source labels alone do not establish target-model difficulty. Missing labels stay uncalibrated; prompt length, rubric count and schema size are not treated as difficulty measurements. See the difficulty audit and the exact rules.

A calibrated revision would record a pinned Qwen checkpoint, multiple independent attempts, outcome-verified success criteria, and per-task pass rates. No such model run is claimed by this release.

MathArena outputs and source-model evidence

arxivmath preserves all 10,420 captured Qwen3.6-35B attempts from the pinned outputs release, grouped into 1,649 distinct questions. An upstream indexing bug repeated one question under several problem indices; 956 questions in the parent training set have no outputs. The outputs source is complete as published, but it does not cover every parent question.

Gold answers are separate from sampled responses and conversations. source_metadata_json.source_reported_baseline and empirical_difficulty retain per-question source-reported attempt counts, correctness counts and pass rates. These describe Qwen3.6-35B, with no claim of independent samples or measured Qwen3.6-27B performance. Each question has 4 or 12 captured attempts; 725 questions have zero source-reported successes and 378 have all attempts reported correct.

Output annotations carry CC-BY-4.0. Matched article URLs, article-license terms and restriction flags remain in source_metadata_json.licensing, source_metadata_json.source_article, and quality_flags. See the MathArena audit.