spacezenmasterr/writer-brief-v3-batch100

Writer Brief v3: dev dataset

Public development dataset (dev dataset), November 2, 2026. All 100 pairs await human review; zero are approved or eligible for SFT. No training has been run on this batch. “Dev dataset” describes the development-stage status of this entire collection. The frozen train/dev/test partitions remain 60/20/20; the dev partition contains 20 documents. The train split name preserves allocation, not training approval.

This dataset pairs an English engineering brief/context C with a complete source document Y. It contains 50 engineering blogs and 50 historical OSS documentation pages, allocated to 60 train / 20 dev / 20 test. Publisher or repository groups do not cross splits. This check does not prove absence of semantic near-duplicates or pretraining contamination.

MiniMax-M3.1-Flash (medium effort) generated the briefs from the full targets. The writer input combines the approved natural brief format with instructions and verbatim technical assets. Each messages value includes system, user, and the complete assistant target. Automated review was a separate model call using the same model; its labels are advisory, not calibrated human quality labels.

Review status

Model verdict	Count
passed	24
needs_revision	75
source_blocked	1

96 review reports satisfy the coverage contract. Four remain incomplete:

• 32a8e4adab10cb15 — Deploy services to a swarm
• 7e5b27a943c0b645 — High availability with Redis Sentinel
• 0d91f0f4d4f0a8c7 — Client-side caching in Redis
• 54dcdca5a8920aaa — Switching user database on a running system | Spotify Engineering

All rows retain human_review_status="pending", eligible_for_sft=false, and training_authorized=false. A model passed verdict does not override these flags. Review source page artifacts, factual/temporal findings, and context sufficiency before any approved training export. Keep dev/test sources excluded from training.

Read the dataset

from datasets import load_dataset
ds = load_dataset("spacezenmasterr/writer-brief-v3-batch100")
example = ds["train"][0]  # review draft only
print(example["context"])
print(example["target"])

This public dataset can be read without an access token.

Field	Meaning
document_id, title, split, document_type	Frozen identity and allocation
brief	Exact generated natural writer brief
context	Exact full user message C, including code/table assets
target	Exact full target Y
messages	System, user, assistant chat sequence
source_url, source_site, source_family, source_license	Source attribution and available license metadata
provenance_json	JSON-encoded full provenance, including pinned OSS revision/path
model_verdict, review_contract_valid, model_review_json	Advisory model review and coverage; detailed review is a JSON string
human_review_status, eligible_for_sft, training_authorized	Pending-only controls
curated_target, *_sha256	Target variant and project content identities
artifact_directory	Repository-relative sidecars

provenance_json and model_review_json are JSON strings so varying source and review schemas load consistently. Decode them with json.loads when needed. Project *_sha256 identities hash the project's canonical JSON serialization, including quoted/escaped strings, rather than raw Markdown bytes. export_manifest.json separately records raw file SHA-256 checksums.

Files and review workflow

• data/{train,dev,test}.jsonl: self-contained Hugging Face rows with C, Y, messages, review, and provenance.
• Review index: download the package and open locally; links are relative.
• Review manifest and summary.
• writing_data/data/writer_brief_v3_batch100_review/drafts/: exact original draft message JSONLs.
• writing_data/data/writer_brief_v3_batch100/ and ..._rechecks/: the 100 selected artifact sets (85 original + 15 recheck results), eight files each.
• writing_data/data/writer_brief_v3_curation_repairs/: Stripe repair proof and repaired full target; its original frozen target is also retained.
• prompts/: exact brief and review prompts, bound by the manifest's prompt hashes.
• docs/: project and pipeline README snapshots. Their original repository-relative code links refer to the source project and are not guaranteed to resolve inside this dataset package.
• export_manifest.json: inventory, raw checksums, validation, and metadata-only portability edits.

Source project: terhuhf/vamos_fire_fire. The local documentation snapshot can be newer than that repository's published HEAD; this dataset does not claim that every local pipeline change has been published. The package excludes raw API calls, runtime logs, credentials, and unrelated corpora.

Target repair and limitations

bcff116005ef1506 (Stripe Railyard) uses a separately versioned repair of five damaged code blocks, checked against canonical Stripe HTML retrieved in November 2026. Original prose and the frozen original target are retained. Historical version scope remains pending review; the repaired Y is not the original frozen Y. The other 99 selected targets are unchanged. No target has been truncated.

Message-text token statistics (system + user + assistant, excluding chat-template overhead): nearest-index p50 5,020; p90 9,536; maximum 22,747. These are preparation statistics, not model quality or training results.

Provenance and rights

No blanket license is assigned to this mixed-source review collection. Source license metadata is retained per row: 50 blog records have no recorded license; OSS metadata lists Apache-2.0 (11), CC-BY-4.0 (13), CC-BY-SA-4.0 (8), MPL-2.0 (8), and MIT (10). Metadata alone does not complete a rights review. Rights to original documents remain with their respective owners; public collection does not grant redistribution rights. Source-by-source rights review remains incomplete. Targets retain original public author/byline information, including source email addresses. Only local machine paths in one repair metadata record were removed; briefs, contexts, targets, and message sequences are preserved exactly.