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 email/creative/general)	~103 words avg, up to 221
total	612	13

Splits

split	rows
train	489
validation	62
test	61

Schema

field	type	description
id	string	Unique task ID (rl_v01_XXXXXX)
family	string	Task family (rewrite_repair, compression, tone_shift, …)
domain	string	Domain (email, slack, technical, …)
mode	string	Task mode (rewrite, direct_generation)
register	string	Target register (casual, terse, warm_professional, …)
instruction	string	Full user-facing instruction
input_text	string	Source text to rewrite (empty for direct_generation)
constraints	dict	Hard constraints (max_words, preserve_numbers, …)
reward_profile	string	Which reward profile to use
required_facts	list[str]	Facts that must appear in a correct response
forbidden_facts	list[str]	Facts that must NOT appear
forbidden_phrases	list[str]	Phrases that must NOT appear
trap_tags	list[str]	Expected failure modes for this task
split	string	train / validation / test
source	string	Dataset version/source

Reward formula

See reward README.

Prime Intellect environment

prime env install jayshah5696/humanize-rl-env
prime eval run jayshah5696/humanize-rl-env --model openrouter/free \
  --api-base-url https://openrouter.ai/api/v1 \
  --api-key-var OPENROUTER_API_KEY \
  --num-examples 20 --rollouts-per-example 4 --max-tokens 1024