asingh15/value-as-tool-gvr-trees

Value-as-Tool branched GVR trees on ArXivMath

Branched Generator → Verifier → Reviser (GVR) rollouts behind the GVR tree explorer Space. There is one tree per problem for 1,649 ArXivMath short-answer problems (1,502 train, 147 eval), generated by Qwen/Qwen3.5-9B.

The model writes a first solution c1. At each of 10 verification points, 4 independent verifier calls judge the current solution, and each verdict routes one revision: says correct → re-check anyway, minor fix → revise, critical flaw → regenerate. The first successful revision becomes the next spine solution, so each tree has a spine c1 → c10 plus three one-step leaves per point. That makes 67,609 candidates and 65,960 verdicts in total. Every candidate is graded by an external answer judge (openai/gpt-oss-20b (answer judge: extracted \boxed{} answer vs gold)).

Layout

• index/trees.json.gz: one row per tree: split, arXiv id, Qwen3.6-35B pass rate and difficulty bucket, spine labels c1–c10, counts of correct candidates and distinct answers, and the problem.
• index/overview.json: dataset-level tables: label changes by revision mode, verdicts by judge label, tree labels by difficulty and spine accuracy by round.
• index/prompts.json: one exact prompt per call kind, with the problem, candidate under review and verifier feedback replaced by ⟦problem⟧, ⟦candidate⟧, ⟦critique⟧ and ⟦category⟧.
• trees/<problem>.json.gz: one tree. nodes are candidates with their call index, parent, cycle, verification point, branch, spine flag, revision mode, answer group, judge label, tokens and text. verdicts hold each verifier call's verdict, fault category and critique, keyed by the candidate they assessed (on). answers lists the distinct extracted \boxed{} answers. calls lists every model call, including forced-recovery turns.
• reasoning/<problem>.json.gz: private reasoning of every call in that tree, keyed by call index.

Files are gzipped JSON. The flat JSONL export (nodes.jsonl, verifications.jsonl, trees.jsonl) is produced by scripts/export_gvr_tree_dataset.py in the value_as_tool repository.

Provenance and terms

Problems and gold answers come from MathArena/arxivmath-training_outputs (CC BY 4.0; MathArena, ETH Zürich SRI Lab), split into train and eval by a hash of the normalized problem text. The difficulty buckets use MathArena's Qwen3.6-35B attempts. Model outputs are from Qwen3.5-9B and are subject to its license. Labels come from an automatic judge (98.7% agreement with MathArena's own grades on 999 Qwen3.6 attempts), so treat individual labels as evidence, not ground truth. Config fingerprint 98c72c0c62c4db27.