# Changelog — Math Corpus

## v2.1.1 — 2026-05-03

**Version:** 2.1.1 · **Previous release:** 2.0.3 (2026-04-19)

Incremental release: new documentation, per-OL solver statistics, and a small amount of additional problems.

### Documentation

- `generation_pipeline.pdf` — overview of the end-to-end problem generation pipeline (graph synthesis → enrichment → text rendering → LLM rewrite → solver verification).
- `dsl_specification.pdf` — formal specification of the CG-Python DSL used to express computation graphs.

### Statistics

- New chart `solve_rate_by_ol.png` — per-`olympiad_level` solve composition for a 9-model capability ladder (top-left = strongest, bottom-right = weakest). Stacked bars show `correct.strict` vs wrong; cells with fewer than 30 attempts are faded as not significant. Embedded in the README inside the LLM Solvers section.

### Data growth (vs 2.0.3)

The corpus grew by **+321 problems** (59,165 → 59,486, +0.5%) — distributed proportionally across domains.

|  | 2.0.3 | 2.1.1 | Δ |
|---|---:|---:|---:|
| Total problems | 59,165 | 59,486 | +321 |
| NT | 39,800 | 39,959 | +159 |
| ALG | 5,685 | 5,774 | +89 |
| COMB | 10,999 | 11,054 | +55 |
| GEOM | 2,681 | 2,699 | +18 |
| Total LLM solver attempts | 233,227 | 252,773 | +19,546 |
| Total strict-correct solutions | 113,366 | 117,945 | +4,579 |

Lemma catalog (88) and template count are unchanged.

## v2.0.3 — 2026-04-19

**Version:** 2.0.3 · **Previous stable release:** 1.1.2.18 (2026-02-15)

### Headline: quality-first release

Release 2.0 reorients the dataset from maximum volume toward per-record quality. Every released problem passes through several quality gates; anything flagged is either dropped or held back for manual review. The corpus shrank from **81,313 → 59,165 problems (−27%)**, but each remaining record is vetted much more rigorously:

- **Only problems solved correctly (strict match) by at least one independent LLM solver are kept.** If no strong model converges to the stored answer, the task is almost certainly a generator bug (bad stored answer, inconsistent statement, missing case) and is excluded.
- **Solver-coverage jump:**
  - mean attempts/task: **1.79 → 3.94** (×2.2)
  - mean correct/task: **0.56 → 1.92** (×3.4)
  - tasks with ≥1 correct solution: **38.6% → 98.9%** (+60 pp)
  - whitelist expanded from **7 → 14 solver models**
- **Skill-inversion check on answer correctness.** For tasks with enough independent solver coverage, the pipeline verifies that the probability of solving the task grows with model ability (θ from an external IRT calibration). When the observed distribution is clearly inverted — weak models succeed while strong ones fail — the task is flagged as suspect: that pattern almost always reflects a bug in the stored answer rather than genuine difficulty. The flagging threshold scales with the amount of evidence available so small-sample flukes do not get tagged.
- **Stronger domain classification** (Phase 2) — context-aware domain intersection; the default domain is now NT. ALG records with strong graph evidence of NT/COMB are correctly reclassified. Less “mislabeled ALG”, cleaner domain assignment overall.

### Continuity with 1.1.2.18

- **~51 k problems** are carried over from the previous release **with the same `id` and unchanged text** (verified by SHA-256).
- **~7.7 k problems** are new, mostly from expanded ALG templates.
- **~29 k problems** from 1.1.2.18 did **not** make it into 2.0 — most were cut by the “≥1 correct solution” gate (many NT problems that no whitelist solver could solve).

### New record fields

- `parent_id` — link to the original record for narrative variants.
- `variant` — variant tag (`narrative` / empty for the mathematical version).

### Format & packaging changes in 2.0.3

- **Parquet output.** `ready.parquet` now ships alongside `ready.jsonl`. Both files carry identical records; `load_dataset` and the HuggingFace viewer use the Parquet copy, while JSONL remains available for direct streaming and `grep`-style inspection.
- **`lemma_applicability` → `list[{lemma, status}]`.** The previous sparse-dict representation is rejected by Arrow-based consumers; the new list-of-struct shape has a fixed schema and works cleanly in HF viewer and `datasets`.
- **Split files removed.** The release no longer ships `train.jsonl` / `valid.jsonl` / `test.jsonl`. Consumers split the single shuffled file themselves — records carry enough structural metadata (`recipe_id`, `seed_template_id`, `lemma_paths`, `olympiad_level`, `domain`) to do structure-aware splits.
- **Per-record `split` field removed.** Without shipped split files there is nothing to pin the label to.
- **Explicit Parquet schema.** Arrow schema is now declared inline by the build pipeline; record drift fails the build loudly instead of being inferred silently.

### Problem generation pipeline

#### Two-stage text: `problem_raw` → `problem`

- The generator writes a deterministic Python-rendered statement into `problem_raw`.
- The final `problem` is produced by an LLM rewriter:
  - **Qwen3-235B-Instruct** — for lighter problems.
  - **gpt-5.1** — for more elaborate ones.

#### Narrative variants

- Eligible problems get an additional record with a narrative (story-style) statement (`{id}_n`, `variant="narrative"`, linked via `parent_id`).
- The release includes **1,320** narrative records (~2% of the corpus).

#### Python renderer (`graph_to_text.py`)

- Deterministic text from the graph — used to produce `problem_raw`.
- Rejects problems with more than 12 `Let`-bindings.
- Fixed precedence for `Mod/Pow` and `-1^k`.
- POLY_ORBIT templates now use `let`-bindings instead of inline nesting (text shrank from ~3,900 to ~500 characters).

### New templates and lemmas

**+21 seed templates** and **+24 lemmas** have been added. The new content is oriented toward **algebra** (QF PSD forms, polynomial orbits, cubic/quartic forms, telescoping sums, systems of equations) and **more elaborate problems** — the mean `olympiad_level` of new templates is 5.5 versus 4.6 over the whole corpus.

Totals in the release: **96 templates, 88 lemmas**.

### Metrics

#### Size and distribution

|  | 1.1.2.18 | 2.0.3 | Δ |
|---|---:|---:|---:|
| Total problems | 81,313 | 59,165 | −22,148 |
| NT | 58,303 | 39,800 | −18,503 |
| ALG | 10,271 | 5,685 | −4,586 |
| COMB | 10,186 | 10,999 | +813 |
| GEOM | 2,553 | 2,681 | +128 |
| Lemmas used | 64 | 88 | +24 |
| Templates | 75 | 96 | +21 |
| Suspects (excluded) | — | 36 | (new) |

The dataset now ships as a single shuffled file. Train/validation/test splits are intentionally left to the consumer so experimental protocols can be chosen per use case.

#### LLM solver coverage (whitelist)

|  | 1.1.2.18 | 2.0.3 | Δ |
|---|---:|---:|---:|
| Whitelist size | 7 models | 14 models | +7 |
| Total attempts | 145,283 | 233,227 | +61% |
| Total strict-correct solutions | 45,378 | 113,366 | +150% |
| Tasks with ≥1 attempt | 87.6% | 100.0% | +12 pp |
| Tasks with ≥1 correct solution | 38.6% | 98.9% | +60 pp |
| Mean attempts per task | 1.79 | 3.94 | ×2.2 |
| Mean correct solutions per task | 0.56 | 1.92 | ×3.4 |