markab/rpn-gpt-data

rpn-gpt data

The training rows and the evaluation sets of rpn-gpt. rpn-gpt puts an exact Forth machine into the weights of NVIDIA Nemotron 3.5 Lightning and trains a LoRA adapter, markab/rpn-gpt-lightning, to use it. This dataset holds the rows that trained that adapter and the sets that measured it.

Files

Config	File	Rows	Size	What it is
rows_r7	train/r7_train.jsonl, train/r7_val.jsonl	9269 / 388	229 MB / 9.0 MB	The rows that trained the round-7 adapter
demos_r5	demos/r5_train.jsonl, demos/r5_val.jsonl	1177 / 53	14.0 MB / 0.6 MB	Round-5 program demonstrations. They are also in rows_r7.
demos_r6	demos/r6_train.jsonl, demos/r6_val.jsonl	534 / 16	6.6 MB / 0.2 MB	Round-6 demonstrations: stack-form programs and recovery after a bug. Round 6 was not promoted, and rows_r7 does not hold them.
frozen_test	eval/frozen_test.json	600	97 KB	The frozen test. Used only at milestones.
dev_round	eval/dev_round.json	407	65 KB	The dev set that decided each round
dev_codefull	eval/dev_codefull.json	200	35 KB	Code-heavy dev set
demo_set	eval/demo_set.json	90	14 KB	Hard problems that need the machine
demo10	eval/demo10.json	10	2 KB	10 hard problems for in-weights runs
dm_frontier	eval/dm_frontier.json	260	31 KB	Easy DeepMind problems, 10 per family
math500	eval/math500.json	500	158 KB	MATH-500, the guard against a plain-math regression
	cards/*.txt		12 KB	The cards: system prompts that describe the machine. reference_v4.txt is the current one. nomachine.txt is the prompt for MATH-500.

An evaluation row has q (the problem), a (the expected answer), family, and for most sets kind (hard, small, variant, easy, easy_big).

A training row is one whole turn as Lightning token ids, ready for SFT:

• input_ids: the turn in the Lightning tokenizer: the card as system prompt, the problem, the model's thinking, its Forth blocks, the machine's output blocks, and the answer.
• train_mask: 1 where the model wrote the token and the loss applies. 0 on the prompt, on the machine's output blocks, and on a masked first attempt in a recovery row.
• q, answer, family, kind, machine (the answer came from the machine), reruns, stage (the source, below), and recovery in demos_r6.

To read a row, decode it with the base model's tokenizer:

from datasets import load_dataset
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16")
row = load_dataset("markab/rpn-gpt-data", "rows_r7", split="validation")[0]
print(tok.decode(row["input_ids"]))

How the rows were made

Every row is a correct, clean turn. Every row has card v4 as its system prompt, also when the rollout used an older card. Every output block is exact against the kernel's token stream. The stage field gives the source:

stage	Rows (train)	Source
A, B, W, WR	2630, 1453, 857, 416	Rollouts of the round-1 adapter on fresh problems (round 3)
R4W, R4M	616, 300	Rollouts of the round-3 adapter on weak families and on the MATH train split (round 4)
demo	1177	Program demonstrations. Code wrote them: a short plan, one program, its exact output, the answer (round 5).
R7	1820	Rollouts of the round-5 adapter: correct, short, answer read from a machine output (round 7)

The rounds, the filters and the results are in training.

Held-out sets

We never trained on an evaluation set. The frozen test was carved out first and was used only at milestones, never to pick a round. Training problems use new seeds. Before this release, we made sure again that no training or demonstration row has the problem text of an evaluation set.

Sources and licenses

Part	Source	License
Problems of the L6.* and DMbig.* families, and the demonstrations	Our generators in rpn-gpt. DMbig uses DeepMind's question forms with bigger numbers.	MIT
Easy problems (arithmetic, algebra, numbers and other DeepMind modules)	Made with DeepMind's mathematics_dataset generator	Generator: Apache-2.0
math500	MATH-500, the 500-problem test subset of MATH from PRM800K	MIT (both)
R4M problems	The MATH train split	MIT
Model text in the rows (thinking, programs, answers)	Rollouts of Nemotron 3.5 Lightning with our adapters	The base model is under OpenMDW-1.1, which puts no restriction on outputs
Output blocks	Our Forth machine (the spec)	MIT

No row holds text from a teacher model. Every row is either a Lightning rollout or a demonstration that our code wrote.

Our MIT license covers our work: the rows as we built them, the demonstrations, the cards, the generator-made problems, and this card. The MATH and MATH-500 problems stay under their MIT license (Copyright (c) 2021 Dan Hendrycks, and Copyright (c) 2023 OpenAI for PRM800K). The DeepMind generator stays under Apache-2.0.

License

MIT, Copyright (c) 2026 Mark Berry and rpn-gpt contributors (LICENSE).

Credit and citation

The MIT license requires that you keep the copyright and permission notice in copies. In addition, we ask you, politely, to credit rpn-gpt with a link to github.com/markab21/rpn-gpt when you use or change this data.

@misc{berry2026rpngpt,
  author       = {Mark Berry},
  title        = {rpn-gpt: An exact Forth machine compiled into the weights of a language model},
  year         = {2026},
  howpublished = {\url{https://github.com/markab21/rpn-gpt}}
}