Datasetpublic

jayshah5696/humanize-rl-sft-dataset

humanize-rl-sft-dataset (v2) 4,835 high-quality SFT pairs for training a model to write natural, direct prose. Part of the humanize-rl project — a two-layer scoring and alignment pipeline for training small models to generate natural, human-sounding text. What this trains A model that can: Write natural Slack messages and emails from scratch. Rewrite stiff/formal/corporate text into direct, human-sounding prose. Fix grammar without making text formal. Shorten and… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-sft-dataset.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes38downloads

Dataset Card

humanize-rl-sft-dataset (v2)

4,835 high-quality SFT pairs for training a model to write natural, direct prose.

Part of the humanize-rl project — a two-layer scoring and alignment pipeline for training small models to generate natural, human-sounding text.

What this trains

A model that can:

—Write natural Slack messages and emails from scratch.
—Rewrite stiff/formal/corporate text into direct, human-sounding prose.
—Fix grammar without making text formal.
—Shorten and compress without losing meaning.

Domain breakdown

Domain	Rows	%
chat	2288	47.3%
email	1453	30.1%
general	800	16.5%
creative	255	5.3%
grammar	26	0.5%
compress	13	0.3%

Quality

—All rows passed a deterministic heuristic quality check.
—All rows passed a Flash Lite LLM judge (naturalness ≥ 4, fact preservation, no AI tells).
—Average naturalness score: 4.91 / 5.0
—Bad-phrase rate (Certainly, Furthermore, etc.): < 0.1%

Sources

Source	Rows
safeexpand3000_raw	1802
stream_b	1617
safeexpandraw	667
chat_expanded	423
curated	326

—safe_expand_*: generated from curated seeds using a safe prompt-based expansion (Arka prompt_based_generator).
—stream_b: instruction-response pairs generated from real human-written text (wardacoder/business-email-dataset, corbt/enron-emails, liamdugan/raid human rows, euclaise/writingprompts).
—chat_expanded: Slack/chat focused pairs generated from hand-crafted direct seeds.
—curated: hand-verified base seeds.

Schema

Field	Type	Description
id	str	Row ID
instruction	str	User task or rewrite request
response	str	Natural human-sounding response
messages	list	ShareGPT format: user/assistant turns
domain	str	chat / email / general / creative / grammar / compress
source	str	Data stream origin
mode	str	directgeneration or rewritehumanize
naturalness_judge	int	Flash Lite judge score (1-5), null for curated rows
version	str	"v2"

Previous version

v1 (1,269 rows) is available as the default config. v2 is a full rebuild with broader domain coverage and stricter quality gates.

Project

This dataset is built and maintained as part of the humanize-rl project.

—GitHub: https://github.com/jayshah5696/humanize-rl
—Goal: Train small open models (Gemma 4 E2B) to natively produce natural, human-sounding text — both from scratch and by rewriting stiff/formal drafts.
—Architecture: Two-layer scoring pipeline (Layer 1 deterministic heuristics + Layer 2 LLM judge), SFT on this dataset, optional RL post-training with DAPO.

License

Apache-2.0. Source datasets have individual licenses — see `data/source_manifest_v04.json` in the training repo for attribution details.