danilxyz/rest-v3

rest-v3

rest-v3 is an English text-rewriting dataset for supervised fine-tuning of a humanizing editor. Each record asks a model to rewrite a source text while preserving its meaning and contains a detector-verified natural-language rewrite.

Dataset composition

The training split contains 1,116 JSONL records:

• —1,033 newly mined, on-policy rewrites from the from-final-best generator checkpoint.
• —83 compatible existing verified examples.
• —541 examples sourced from the API corpus and 575 from the web corpus.

Every record includes chat-formatted messages, the source and winning rewrite, a normalized source hash, source type, quality metadata, GPTZero verdict, register features, token-edit distance, and a normalized training weight.

Construction

The new examples were built in two stages:

1. 1.Phase 0 — generation and quality gates. A balanced corpus of 1,362 English texts (681 API and 681 web) produced eight mode-diverse rewrites per source. All candidates passed semantic, length, and faithfulness gates before selection. Candidates with a token-edit ratio below 0.20 were excluded from detector verification.
2. 2.Phase 1 — detector verification. Register-eligible candidates were ranked by proximity to the target register and checked with GPTZero. Verification stopped for a source after its first human verdict. Across 1,822 checks, 1,033 rewrites received a human verdict.

The verification run used 1,031,053 detector words. No model training was performed as part of dataset construction.

Weights and legacy data

For each record with complete metadata, the raw training weight is difficulty_weight × register_proximity. difficulty_weight is inverse quality-pass frequency across K=8 candidates, and register_proximity measures closeness to the target function-word, adjective, and lexical-diversity ranges. Source-type scaling then equalizes the mean weight between API and web records, preserving the dataset's observed source-type count ratio instead of allowing a harder source type to dominate total loss mass.

The 17 Turso-only historical examples without weight metadata are excluded. The recovered gate-fix example is merged when its source hash is not already represented; duplicate source hashes resolve to the newly verified Phase 1 example.

Intended use

Use this dataset for research and supervised fine-tuning of English text-to-text rewriting systems. The difficulty_weight, quality_pass_fraction, and register metadata support weighted or stratified training analyses.

Limitations

GPTZero human verdicts are detector outcomes, not proof of human authorship or universal quality. The data is synthetic and inherits limitations from its source corpora, generation checkpoint, quality gates, and detector. Review suitability, licensing, privacy, and safety for your deployment before use.