README.md

---
dataset_info:
  config_name: SODA_2026
  features:
  - name: question_id
    dtype: string
  - name: paper_id
    dtype: string
  - name: title
    dtype: string
  - name: venue
    dtype: string
  - name: year
    dtype: int64
  - name: node_id
    dtype: string
  - name: prompt_type
    dtype: string
  - name: problem
    dtype: string
  - name: answer
    dtype: string
  - name: answer_type
    dtype: string
  - name: answer_symbols
    dtype: string
  - name: answer_assumptions
    dtype: string
  - name: reference_solution
    dtype: string
  - name: difficulty
    dtype: string
  - name: topic_tags
    dtype: string
  - name: format_issues
    dtype: string
  - name: verification_method
    dtype: string
  - name: procedural_audit
    dtype: string
  - name: content_audit
    dtype: string
  - name: independent_verification
    dtype: string
  - name: equivalence_check
    dtype: string
  - name: source_dag_repo
    dtype: string
  - name: source_dag_config
    dtype: string
  - name: source_dag_split
    dtype: string
  - name: source_dag_revision
    dtype: string
  - name: source_dag_sha256
    dtype: string
  - name: generation_model
    dtype: string
  - name: generation_effort
    dtype: string
  - name: generation_prompt_version
    dtype: string
  - name: reused_from_four_question_run
    dtype: bool
  splits:
  - name: test
    num_bytes: 1165731
    num_examples: 290
  download_size: 545335
  dataset_size: 1165731
configs:
- config_name: SODA_2026
  data_files:
  - split: test
    path: SODA_2026/test-*
---

# TCSBench 2026: Final-answer problems

This release contains **290 exact-answer problems** from **142 SODA 2026 paper DAGs**. It is SODA-only, and answers are visible. The public config/split is `SODA_2026` / `test`.

## Coverage and curation

The generator requested three problems per DAG, each with a distinct theorem-like target node; quality gates did not force-fill unsuitable papers. After curation, 62 papers have three questions, 38 have two, 28 have one, and 14 have none. Thus 128 papers appear in the question rows. [`paper_coverage.json`](./paper_coverage.json) lists all source papers and their final counts.

This release was copied from the [original AI-Math-TCS release](https://huggingface.co/datasets/AI-Math-TCS/TCSBench_2026_ans), then the trailing answer-format instruction was removed from **every** public `problem`. The mathematical question, canonical answer, reference solution, audit fields, and `question_id` were otherwise preserved. **41 rows were excluded** because their canonical `answer` contains an integer literal of six or more decimal digits anywhere in the expression, including inside fractions or radicals. This is a textual literal rule, not an evaluation of symbolic expressions. The original source release remains available; this copy is the curated evaluation set.

Remaining answer types: 133 integer, 86 rational, 41 rational-expression, and 30 algebraic-expression. 1 retained row(s) have a `unicode_math` formatting flag, recorded in `format_issues`.

## Generation and checks

Problems were generated with `gpt-5.6-sol` at `xhigh` effort from a target node and its proof-relevant DAG context. Each accepted problem passed context and exact-answer eligibility checks, scaffold reduction, a proceduralness audit, an independent solve, answer-equivalence review, and a final content audit. The post-generation removal of the response-format instruction changes no mathematical assumptions or requested quantity, but these audits were not rerun on the edited string. Model-assisted audits are not a formal guarantee of mathematical correctness.

The `answer` column is the canonical closed-form answer; `reference_solution` explains it. `answer_symbols` and `answer_assumptions` are JSON strings used by the 0/1 grader. Exact symbolic comparison is attempted first, with `gpt-5.6-luna` at medium effort used only when exact comparison is unresolved. See the [grader](https://github.com/CohenQU/AI-Math-TCS/blob/6c1dc00/scripts/synthetic/final_answer_grader.py) and [equivalence fallback](https://github.com/CohenQU/AI-Math-TCS/blob/6c1dc00/scripts/synthetic/final_answer_equivalence.py).

## Provenance

Each row records its source paper and target node, generation settings, audit records, and the SHA-256 of the exact source DAG payload. The source is the [structurally validated SODA 2026 DAG split](https://huggingface.co/datasets/AI-Math-TCS/tcs_dags/viewer/SODA_2026/legacy_v1_6_structurally_repaired_gemini_3_8_flash) at revision `4ccf1e006ac1d6f3053049b7fcc5aa47cec8a6f8`.

```python
from datasets import load_dataset

ds = load_dataset("HerrHruby/TCSBench_2026_ans", "SODA_2026", split="test")
```