humanize-rl
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.humanize-rl-tasks-v03
humanize-rl RL Tasks v03
492 RL tasks designed to train and evaluate models that produce human-sounding prose
under specific constraints. Each row pairs an authored instruction with a constraint
spec consumable by the humanize-rl reward
environment (deterministic checks + Layer-1 heuristics + ridge regression style scorer).
Built per docs/plans/v03-rl-tasks-dataset.md.
What's in a row
field
meaning
id
stable task id (rl_v03_NNNNNN)
family
top-level… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-tasks-v03.humanize-rl-prime-sft-messages-env0314
Humanize-RL Prime SFT Messages Env0314
Prime prime-rl SFT dataset for Humanize-RL.
Schema: each row has a messages list with one user instruction and one assistant target.
Splits:
train: 4313
validation: 239
test: 241
total accepted: 4793
rejected upstream by builder: 62
duplicate ids across published splits: 0
repair-reference rows: 20
Source artifact: v04_sft_final_plus_llama_failure_refs_env0314, built from restored v04 SFT data plus the clean Llama failure-reference repair… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-prime-sft-messages-env0314.humanize-rl-research-artifacts-env0314
Humanize-RL Research Artifacts Env0314
This dataset archives the ignored local artifacts needed to continue the Humanize-RL reward patch, SFT repair, and Qwen 2B/9B ablation work after leaving the original worktree.
Primary source run: zztqgqclh3y3hslpjsofzpcf
Prime env: jayshah5696/humanize-rl-env@0.3.14
W&B run: https://wandb.ai/jayshah5696/humanize-rl/runs/akzopsz9
Published SFT dataset: jayshah5696/humanize-rl-prime-sft-messages-env0314
Contents… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-research-artifacts-env0314.humanize-rl-prime-sft-messages-env0315-clean50-primecompat
jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50-primecompat
Prime prime-rl supervised fine-tuning dataset for Humanize-RL.
This is the env0315_clean50 S2 repair-data candidate. It starts from the
env0314 Prime SFT corpus and adds cleaned env0315 repair references generated
from saved Prime rollout-audit failures.
Splits
split
rows
train
4358
validation
242
test
243
total
4843
Sources
source
rows… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50-primecompat.humanize-rl-tasks
humanize-rl-tasks
Single-turn RL task dataset for the humanize-rl project.
Each row is one writing task. A model receives the prompt (instruction + source text),
produces a completion, and the environment scores it with the 50/50 reward formula:
reward = 0.50 × ridge_rubric_mean + 0.50 × deterministic_mean + penalties
Dataset composition
source
tasks
families
input length
v01_template
100
10 (template-generated)
~20 words
v02_real_source
512
3 (real… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-tasks.