tcs-dags-clean

FOCS 2024 paper dependency DAGs with every depends_on edge audited, and the standalone proof questions that target them, as used by the tcs knowledge-injection experiments. Private working copy; not the canonical AI-Math-TCS release. Built by scripts/hf_export.py in the tcs repo.

from datasets import load_dataset
dags = load_dataset("3N3G/tcs-dags-clean", "dags", split="FOCS_2024")
questions = load_dataset("3N3G/tcs-dags-clean", "questions", split="FOCS_2024")

venue	papers	questions	questions kept
FOCS 2024	91	2599	2053

Sources

• DAGs: AI-Math-TCS/tcs_dags FOCS_2024 split gpt_5_4 at revision 24965d89a53f88d7caa9ee5fd7abf8481f11f452, with the edge audit below applied.
• Questions: AI-Math-TCS/tcs_standalone_hard FOCS_2024 gpt_5_4 at revision 15bea05f1ae30580bf98aa8e6c5b9871a18f0ed6, all rows.
• arXiv ids and v1 dates: resolved by title match (exp1.9); arxiv_match is matched or unresolved.

STOC 2026 is deliberately absent: its upstream graphs are not yet audited this way.

Config dags

One row per paper: venue, paper_id, title, authors, year, arxiv_id, arxiv_v1_date, arxiv_match, edge_repaired, num_nodes, num_edges, dag (JSON: nodes, edges, deleted_depends_on_edges, dropped_depends_on_edges). Edge convention: from requires to.

The edge audit. Upstream's FOCS 2024 graphs are not acyclic: 90 of 91 papers have directed cycles on depends_on alone, 37.6% of nodes inside nontrivial strongly connected components, and 9.3% of the edges outside cycles run from a definition or cited fact to a result proved in the paper. Every one of the 9,828 depends_on edges was judged keep / flip / delete by gemini-3.1-pro-preview (thinking high) from node statements, summaries and the edge's own description, under the prompt proposed as Stage 2.75 of the upstream pipeline (data/edge_repair/audit_prompt.txt in the tcs repo); residual cycles got one targeted second pass. Result: 6,840 kept, 1,303 flipped, 1,685 deleted; every paper acyclic on depends_on. Every surviving depends_on edge carries provenance (kept | flipped) and the auditor's audit_reason; deletions sit under deleted_depends_on_edges with reasons. Other relations (elaborates, proves, ...) are upstream's verbatim with provenance: original and were not audited.

Validation against an independent cycle-only audit of the 3,909 intra-cycle edges: 95.6% identical verdicts, 96.6% agreement on whether the recorded direction stands, all four hand-adjudicated edges reproduced. Known residue: 11 kept edges still run from a definition or cited fact to a derived result. Per-edge correctness beyond that is verified only by sampling; the reasons are in the data for review.

Config questions

One row per generated question, filtered rows included: question_id (venue/paper_id/node_id/row_idx), venue, paper_id, node_id (zero-padded), difficulty, topic_tags, problem, answer, rubric, proof_source, generated_reference (always null here), kept, filter_reason. kept applies the experiments' loader rules: proof_source_not_dag (reference answer is model-written, not the paper's proof), leaked_generator_text (generation-guideline prose shipped inside the problem), node_missing (target node absent from the DAG), no_extracted_proof (target has no proof or sketch in the DAG).

Known defects

• Standalone problems restate their prerequisites as givens, so they cannot serve as a bare theorem statement.
• Larger nodes' proof_content often holds a whole section, embedding smaller nodes' proofs verbatim (118 cases); any context built from other nodes' proofs must screen for this.
• Node summaries frequently describe the proof plan.
• Close reading of sampled questions found a provably false statement and an undefined term.
• Some proof_source=dag rows target cited results whose "proof" is citation text.