KinyoubiAtelier/tulana-items-classf

Tulana benchmark: studio-original items, Class F (Pool B)

Status: experimental, item content approved, license declared, renderings PROVISIONAL — see RELEASE_LABELS.md and "Approval status" below. Maintainer: Kinyoubi Atelier & Co. (@ankitkinyoubi) Last updated: 2026-09-02 Publication state: published at KinyoubiAtelier/tulana-items-classf — first revision 2026-09-02, released on the founder's explicit per-artifact instruction (program charter standing rule 1 satisfied for this artifact; license CC-BY-4.0 declared the same day under rulebook F-2). Renderings remain PROVISIONAL pending founder review (see Lifecycle); numbers derived from them keep the PROVISIONAL label.

---

What this is

48 studio-original instruction-following items in native Odia (batch 001), each with one or two machine-readable constraints (a kinyoubi.* checker), an English gloss, and per-item provenance. This is Tulana's Pool B — items the studio authored itself, as opposed to Pool A (KinyoubiAtelier/tulana-bench-anchor, staged, not yet published), which is redistributed IndicIFEval data. This repo also ships the pilot's constructed renderings of these 48 items: a C1 (native) projection plus C2 (romanized, canon + a seeded variant) and C3 (code-mixed, 25/50/75% intensity) transformations, with full per-rendering provenance and replacement logs.

Approval status (read carefully — two separate gates, not one)

1. Item content — approved. The founder recorded a blanket verdict on 2026-09-02: "ALL 48 ITEMS APPROVED" (data/REVIEW-QUEUE-batch-001.md), after reviewing the digit-normalized batch. No edits were requested. Every item's status field now reads founder-approved-2026-09-02, and this is the founder's own recorded final verdict — not a lab inference from silence.

2. License — declared. The founder declared CC-BY-4.0 for this artifact on 2026-09-02, per provenance rulebook rule F-2 (studio-chosen license, declared per artifact). See LICENSE-NOTICE.md and LICENSE; every per-item provenance record in data/classF-v0-batch-001.provenance.jsonl carries the declaration.

3. Renderings — still open. Approval of item content and declaration of a license do not extend to the C2/C3 renderings: the renderings in data/renderings/ are machine-draft and have not themselves been founder-reviewed. Item-content approval and rendering approval are the spec's two distinct gates (BENCHMARK-SPEC-v0 §2.2/§2.3; spec addendum D8): a correct native item can still have a flawed romanization or code-mix construction. Every number computed from these renderings remains PROVISIONAL until that separate review is recorded (per rulebook S-4, the renderings themselves inherit the now-declared CC-BY-4.0 once that review clears — the license question is settled even though the content-quality question is not).

Why this repo is separate from the anchor pool

This repo carries CC-BY-4.0; the anchor pool is staged under the same license as KinyoubiAtelier/tulana-bench-anchor (not yet published on the Hub — that link will resolve once its own founder checkpoint clears), so the split is no longer driven by a license mismatch — it is driven by pool and provenance-chain separation: this repo is studio-original content (Pool B) with its own review lifecycle (item approval and license declared; renderings still pending review), while the anchor repo is redistributed third-party data (Pool A) with a wholly different provenance chain and attribution obligation. The spec's pool-separation rule (BENCHMARK-SPEC-v0 §8.1: Pools A and B are never merged in a single number) makes keeping their source repos separate the natural structural choice, independent of the license question that originally forced it.

PROVISIONAL labeling (binding, do not drop)

Every number computed from this data in any Tulana release — cell scores, deltas, audit rates — carries the label PROVISIONAL in the table cell or column header, not only in a footnote, for as long as any input to that number (a rendering, a not-yet- declared license) remains unreviewed. Concretely, as of this staging:

• Numbers from the C1 (native) items: item content is approved; PROVISIONAL still applies only insofar as this pool has not yet cleared the anchor-comparability restrictions that apply program-wide (spec addendum D1) — not because of open item review.
• Numbers from the C2/C3 renderings: PROVISIONAL because the renderings themselves are machine-draft, unreviewed (see "Approval status" above).

Renderings: what exists and what doesn't

Suite	Items	Eligibility
renderings/classF-c1-suite.jsonl	48	C1 projection of every approved item
renderings/classF-c2-suite.jsonl	138	C2-canon + C2-var(k=2) for every eligible item
renderings/classF-c3-suite.jsonl	138	C3 @ 25/50/75% for every eligible item
renderings/provenance.jsonl	276	one rulebook §4 provenance record per rendering
renderings/render_log.json	—	full per-item substitution/replacement/transform log

46 of 48 items are renderable; 2 (tulana-f-011, tulana-f-015) are C1-only — every constraint on those two items is excluded under transformation per spec §3, so no C2/C3 rendering exists for them. This is stated in renderings/manifest.json and repeated here rather than left for a reader to notice by absence.

All renderings are labeled constructed, not observed (rulebook S-1) in every record; no ecological-validity claim is made about how anyone actually writes romanized or code-mixed Odia (rulebook S-2). Per rulebook rule S-4, every rendering inherits its parent item's license — CC-BY-4.0, now that batch 001's license is declared (see above) — independent of the still-pending content review of the renderings themselves.

Anchor duplication check

Every item was compared against all 858 pinned IndicIFEval Odia rows (the staged tulana-bench-anchor pool) by character-5-gram Jaccard similarity over the NFC-normalized prompt text. Batch maximum similarity to any anchor row: 0.214 (tulana-f-030); batch mean of per-item maxima: 0.102 — no duplicates or near-duplicates found.

Constraint distribution (batch 001: 48 items, 72 constraints)

Family	Constraints	Share
keyword	23	32%
length	21	29%
format	19	26%
structure	9	12%

All 15 registered kinyoubi.* checkers are exercised at least once.

Structure

data/
  classF-v0-batch-001.jsonl              48 items: prompt, constraints, English gloss,
                                          LLM-assistance disclosure, status
                                          (founder-approved-2026-09-02)
  classF-v0-batch-001.provenance.jsonl   one Class F provenance record per item
                                          (validation field now filled per the
                                          founder's recorded verdict)
  REVIEW-QUEUE-batch-001.md              the founder's review document, including the
                                          2026-09-02 blanket-approval verdict banner
  renderings/
    classF-c1-suite.jsonl                48 native-condition items (runnable suite shape)
    classF-c2-suite.jsonl                138 romanized renderings (canon + var)
    classF-c3-suite.jsonl                138 code-mixed renderings (25/50/75%)
    manifest.json                        pipeline commit, seed, config hashes, counts
    provenance.jsonl                     276 per-rendering provenance records
    render_log.json                      full substitution/replacement/transform log

Lifecycle

1. Founder works through data/REVIEW-QUEUE-batch-001.md — done for item content (2026-09-02, blanket approval).
2. Founder reviews the renderings in data/renderings/ — pending. Corrections force a re-render of the affected item before the PROVISIONAL label on numbers using it comes off.
3. Founder declares a license (rulebook F-2) — done: CC-BY-4.0, 2026-09-02.
4. Publication of anything derived from this pool remains gated on the founder's explicit per-artifact approval. This repository authorizes no publication on its own.

Citation

See CITATION.cff.

Contributing / Security

Staged, not yet public. See KinyoubiAtelier/tulana-eval's CONTRIBUTING.md / SECURITY.md as the program-wide template once this repo is live.

License

CC-BY-4.0, declared by the founder 2026-09-02 (provenance rulebook rule F-2) — see LICENSE-NOTICE.md and LICENSE. This still ships as a repo separate from KinyoubiAtelier/tulana-bench-anchor (see "Why this repo is separate" above) — the pool-separation rule, not a license mismatch, is what keeps them apart now that both carry CC-BY-4.0. The renderings in data/renderings/ are licensed the same way (S-4 inheritance) but remain content-unreviewed — see "Approval status."

ସମୀକ୍ଷାରେ ସାହାଯ୍ୟ କରନ୍ତୁ · Help review

ଏହି ଡାଟାସେଟ୍‌ର ରୋମାନ୍ ଓ ମିଶ୍ରିତ ରୂପାନ୍ତରଗୁଡ଼ିକୁ ଏପର୍ଯ୍ୟନ୍ତ କୌଣସି ଓଡ଼ିଆ ପାଠକ ଯାଞ୍ଚ କରିନାହାନ୍ତି। ସମୀକ୍ଷା କିପରି ଚାଲେ, ପାରିଶ୍ରମିକ ସର୍ତ୍ତ ଏବଂ ଗୋପନୀୟତା ସୂଚନା ଏଠାରେ: https://huggingface.co/spaces/KinyoubiAtelier/tulana-review

The romanized and code-mixed renderings in this dataset have not yet been checked by Odia readers. How the review works, the compensation terms and the privacy notice are on the review page: https://huggingface.co/spaces/KinyoubiAtelier/tulana-review (registration opens shortly).