Mahesh111000/k-program-environment

K Program Environment (KPE)

Program synthesis in a 12-primitive subset of the K array language. Each problem gives an input array x and a target; the model writes one K expression that maps x to the target. A problem counts as solved only if the expression reproduces the displayed pair and all 8 hidden input-target pairs.

Splits

split	problems	depth	use
train	2,900	2: 1,100 · 3: 1,100 · 4: 700	RL training (Card+Facts scaffold, one attempt)
eval	220	1: 20 · 2: 100 · 3: 100	held-out problems behind every reported K number
eval_depth4	100	4: 100	held out, not in the reported numbers (base accuracy is at the floor)

The training and evaluation problems are disjoint: no shared id, no shared displayed input-target pair, and no training reference program equals an evaluation reference program.

Fields

• id: problem id.
• depth: number of composed operations in the reference program.
• ref: one reference solution (every reference reproduces its displayed pair and all hidden pairs under kmini.py).
• x, target: the displayed input and target. Every target is stored as a list; a single-number target t is stored as [t] with target_is_scalar = true.
• target_is_scalar: whether the target (and every hidden target) is a single number rather than a list.
• hidden: 8 further [input, target] pairs used only for scoring, stored with the same convention.

The functions in k_prompts.py take a dataset row directly; to_task(row) restores the task format of the environment (scalar targets as plain numbers).

Prompts and scoring

k_prompts.py reproduces the served prompts exactly and scores answers; kmini.py is the interpreter used for every training reward and every evaluation.

from datasets import load_dataset
from k_prompts import render_eval, render_train, extract_answer, is_correct

ds = load_dataset("<this repo>")
t = ds["eval"][0]
prompt = render_eval(t, "card_facts")   # "no_card", "card", "facts", "card_facts", "paraphrase"
train_prompt = render_train(t)          # Card+Facts, one attempt (the training prompt)
# score = is_correct(extract_answer(model_response), t)

• Scaffolds. No-card shows only the task; Card adds the operator reference card; Facts adds up to three computed facts about the target; Card+Facts adds both. Paraphrase rewords every sentence of Card+Facts and serves its blocks in reverse order; it is used only at evaluation.
• Attempts. Training uses a single attempt. The evaluation prompts allow up to three attempts with the interpreter's output on the displayed input shown after each; the reported numbers score only the first attempt.