hyz-is/arandu-v1

Arandu v1

One dataset, fused from the fifteen packages that were prepared separately. It exists to teach a model our own framework — Arandu, Hesape, Kyse, Aru and the modules around them — from our own source, at pinned revisions.

Read this first: 102 of 8 473

Only 102 records are admitted for training. The other 8 371 are proposals that nobody has reviewed.

training_eligible is the owner's decision, per the project's twelfth rule, and no step of this pipeline sets it. The 102 inherited it from the component package whose owner recorded the decision. Everything else defaults to false, which is what "not yet judged" looks like.

The admitted config gives you exactly those 102. If you train on anything else, you are training on unreviewed proposals — mostly source reconstruction exercises generated from our files and never executed, so nothing attests that the answer compiles.

config	records	what it is
admitted	102	training_eligible: true. The reviewed material.
train	122	what the components placed in train/. 102 admitted, 20 not.
review	8 339	proposals awaiting judgement. Not training data.
heldout	6	reserved for evaluation. Never train on these.
validation	6	reserved for evaluation. Never train on these.

Licence: there is no blanket grant

This dataset is not uniformly licensed, and the license: other above is deliberate. Licence is per artifact.

• 11 293 records are our own material: licence.artifact: "proprietario", licence.commercial: "permitido".
• 246 records publish licence.commercial: "a_conferir" — the question is open, not answered yes. Two reasons, both recorded per record:
  • the component's own card refuses a blanket licence in writing. htmx-go-v1 says "per source; no blanket license granted for mixed artifacts"; four others say the equivalent in rights. That declaration is carried into licence.artifact verbatim rather than overwritten.
  • the record carries third-party provenance. The 44 htmx-go-v1 records reproduce upstream htmx (0BSD) and Go standard library (BSD-3-Clause) source; the upstream licence is in licence.third_party and the per-source digests in origin.commit_sources.

licenses/ holds 114 licence files, each prefixed by the component it came from. Third-party material under a different regime — arandu-quality-v3/external/, 1 100 Tupi-Guarani records — is excluded from this release, not relicensed.

The fifteen components

Every record carries origin.dataset naming the component it came from and origin.path naming the file inside it.

component	representative	admitted	identical copies dropped	families
arandu-coder-v2	5 455	0	0	17
arandu-coder-v1	2 660	0	2 551	13
arandu-quality-v3	154	50	50	40
ybyra-organization-v0	32	0	0	32
curriculum-completion-v1	30	0	30	30
arandu-deep-v2	29	11	130	21
practical-delivery-v1	24	0	24	6
htmx-go-v1	22	22	22	22
arandu-security-v1	19	19	19	19
agent-project-protocol-v1	12	0	12	6
kyse-design-v1	12	0	0	10
mercado-nacional-v0	12	0	0	12
security-review-qa-v1	12	0	0	3
ayra-native-v1	0	0	228	0
common-style-v1	0	0	0	0

Two components contribute no record of their own:

• ayra-native-v1 is contained entirely in arandu-coder-v2. All 228 of its records have a byte-identical twin there, because ayra is one of the nineteen repositories arandu-coder-v2 was generated from. Nothing was lost — the 228 are in duplicates.parquet with duplicate_of naming the survivor.
• common-style-v1 holds no example. It is a skills package: sources and style rules, no instruction records.

arandu-coder-v1 is superseded by arandu-coder-v2, which regenerated the same corpus at current revisions. 2 551 of v1's records are byte-identical to a v2 record and were passed over. The remaining 2 660 differ because the source moved between generations; most have a near-duplicate in v2. Check near_duplicate_count before using them.

How duplicates were handled

Nothing was deleted. 11 539 records go in and 11 539 come out. Three thousand and thirty-three groups of byte-identical content were found, plus 23 912 near pairs above 0.75 Jaccard similarity.

Within a group of identical content one record is marked representative: true by a written rule, applied in this order:

1. an admitted record wins — dropping the owner's decision is not the pipeline's to make
2. otherwise the later origin.packaged_at wins — order, not a judgement of quality
3. otherwise reserved before train before review, so a held-out example is never demoted in favour of a training copy
4. otherwise the lowest id, so two runs choose the same

Every passed-over copy — 3 066 of them, including the 169 that were the same record filed in two places inside one component — stays in duplicates.parquet with duplicate_of naming the survivor.

Near duplicates are not removed: similarity is not identity, and two exercises over the same file legitimately resemble each other. They are counted per record in near_duplicate_count / near_duplicate_datasets, and every pair is in duplicate-groups.jsonl.

Record shape

{
  "id": "arandu-security-v1/BP-E01",
  "role": "coder",
  "family": "named-route-contracts",
  "split": "train",
  "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
  "origin": {"dataset": "...", "revision": "...", "record_id": "...",
             "repository": "...", "commit": "...", "commit_sources": [],
             "path": "train/coder.jsonl", "packaged_at": "2026-09-08T22:56:30Z",
             "split_declared": ""},
  "licence": {"artifact": "...", "origin": "...", "third_party": [],
              "commercial": "permitido", "declared": true},
  "training_eligible": true,
  "status": "independently_source_reviewed",
  "content_hash": "...",
  "representative": true,
  "duplicate_of": "",
  "near_duplicate_count": 0,
  "near_duplicate_datasets": []
}

role is normalised to lower case — the components wrote Orc and orc for the same role, which split the column in half. origin.split_declared keeps the component's own split string when it is not one of ours: security-review-qa-v1 says training_candidate_family_unassigned, which is real information about that record and is preserved rather than replaced.

origin.revision identifies the component's packaging. Where the component's manifest records a sha256, that is used. Where it records none — two older packages do not — the revision is a content-digest: over the manifest plus the bytes of every example file, so it changes when the content changes. An earlier version digested only the manifest, which for ybyra-organization-v0 holds counts and no digests at all: editing an example without changing a count returned the same revision, and a revision that does not move when the package moves traces nothing.

family is what holds a split together: variants of one task must not straddle train and evaluation. It is read from family, then family_id, then area, then the first source_ids / source_refs entry — always a field the component actually wrote, never invented.

The serialized release: releases/arandu-admitted-v1/

The admitted config, prepared for SFT. It is the 102 records tokenized against the pinned Qwen/Qwen3.8-27B tokenizer at revision 1d4bf0f2, with the prompt masked to -100 and only the assistant turn supervised.

examples	102
families	78
total tokens	269 253
supervised tokens	14 910
shortest / median / longest	743 / 1 614 / 28 426
over 8 192 tokens	3
answer roundtrip verified	102 / 102
truncated	0

Rows are ordered by ascending sequence length, so a length ruler can be walked without re-sorting. No packing, no padding, no synthetic reasoning.

The serializer was validated before it was used here. Run against the earlier first-pilot-v1 release it reproduced token counts and the rendered sha256 on 19 of 19 examples; only then was it applied to these 102.

model/ is not duplicated inside the release — the tokenizer is identified by revision and per-file sha256 in serialized/serialization.json, and the component notices are in licenses/ at the repository root.

Two limits are declared rather than resolved, in release.txt:

• Length is not a memory admission. Three examples exceed 16 384 tokens and one reaches 28 426. On 16 GB T10s that is what decides whether a recipe fits, and it was not measured — no GPU was touched by this preparation.
• No common system turn was added. Eighty of the 102 carry no system message and twenty-two carry their own. The earlier pilot added a context written for patch-proposal exercises; these 102 include explanations too, so applying that frame would assert something the records do not say.

Nothing here ran a gradient, loaded weights, or touched hardware.

Files beyond the splits

• duplicates.parquet — 3 066 passed-over copies, each naming its survivor
• duplicate-groups.jsonl — every exact group and every near pair, with similarity
• reservations/ — 576 source files the components reserved out of training, by repository and path. Published as provenance so the reservation is auditable.
• manifest.json — the fusion record: component counts, the rules above
• components/*.manifest.json — the original manifest of each of the fifteen packages
• licenses/ — 114 licence files, prefixed by component

What this dataset does not contain

• Execution evidence. The reconstruction and completion exercises are proposals generated from source. No record is backed by a run that succeeded. status says so per record.
• A balanced distribution. The corpus follows repository size. hesape dominates the arandu-coder-* components; kyse is a rounding error in them. The imbalance is recorded, not corrected.
• External material. arandu-quality-v3/external/ is excluded.

Reproducing

python3 cluster/runtime/corpus.py ingest --package <component> --out ingest/<component>.jsonl
python3 cluster/runtime/corpus.py group  --in ingest/*.jsonl --out grupos.jsonl
python3 cluster/runtime/corpus.py fuse   --in ingest/*.jsonl --groups grupos.jsonl --out acervo.jsonl