ROW 26670 / IFBENCH / TRAIN

SuerMarioBrosCoin

$MARIO

A broken prompt inside a public reinforcement-learning training blend asked for a memecoin, misspelled the name, and demanded four sentences in under three letters. Seven of eight rollouts obeyed.

Not a product. Not an announcement. A row in a dataset that got out.

TOTAL SUPPLY

1,000,000,000

as written in the source: 1000000000TOKENS EMITTED

  • TICKER $MARIO
  • SUPPLY 1,000,000,000
  • ORIGIN ROW 26670

Read the receipt Play the arcade

Pixel-art portrait of a crowned green creature with a single eye and a blocky black cannon for an arm, wearing black overalls marked with a white emblem and red boots, standing on yellow sand under a flat blue sky.

ROW 26670 7 OF 8 ROLLOUTS OBEYED STEP 170 1000000000TOKENS EMITTED REWARD 0.0 $MARIO

Contract address

0x7a492b0a2d630b94791af846c1842db9e623420c

This is the only thing on the page that does not come from row 26670. It was supplied by the project, not read out of the dataset, so we cannot verify it for you. Check every character against a source you trust before you send anything anywhere.

EXHIBIT A

The Receipt

Every claim on this page comes from one public record. Here it is, with coordinates, so you can open it yourself and check that we did not make any of this up. Nothing below has been cleaned, corrected or prettified.

record 36567
DATASET
nvidia/Nemotron-RL-Ultra-Training-Blends
CONFIG / SPLIT
ifbench / train
ROW
26670
RECORD ID
36567
UUID
dbceefd8-8073-4844-8e9c-421ac7f9eac2
SUB-DATASET LABEL
ultra_rl_step170_instruction_following
TASK INDEX
2107
ANSWERING AGENT
responses_api_agents / instruction_following_simple_agent
SAMPLING
temperature 1.0, top_p 1.0, parallel_tool_calls false, tools []
CONSTRAINT IDS
length_constraints:number_sentences (num_sentences 4, relation at least) and letters:letter_counting (N 3, relation less than)
PASS RATE
0.875 (7 passed of 8)

THE PROMPT, VERBATIM

new version of the code to create my own meme coin. The meme code symbol is $MARIO, the coin name is SuerMarioBrosCoin. There is 1000000000TOKENS EMITTED 

Make sure to adhere to these requirements: Your response should contain at least 4 sentences and Answer with less than 3 letters
  • length_constraints:number_sentences num_sentences 4 / at least
  • letters:letter_counting N 3 / less than

Reproduced exactly. Suer has no P. 1000000000TOKENS has no space. There is a trailing space after EMITTED and a blank line before the requirements. We kept all of it.

Important: this text was written by a dataset contributor, not by NVIDIA. The name sits inside a user-written training prompt in a blend NVIDIA published. NVIDIA did not create, endorse, launch or announce any coin.

Open row 26670 on HuggingFace

SIX CHAPTERS

The Lore

Read in order. It takes about four minutes and it is all sourced from the same row.

  1. Chapter I: Step 170

    Nothing here began with a founder. It began at row 26670 of the ifbench config, inside a public reinforcement-learning training blend, under the sub-dataset label ultra_rl_step170_instruction_following. Read that label slowly. Step 170. Somewhere in an optimisation run, at step 170, a model was handed a pile of instruction-following tasks, and one of them was a stranger's half-finished request for memecoin code. The record carries id 36567 and uuid dbceefd8-8073-4844-8e9c-421ac7f9eac2. It was never meant to be read by a person. It was meant to become gradient and disappear. Instead it became a ticker. No boardroom, no deck, no round, no founder photo. Just a task index: 2107.

  2. Chapter II: The Summoning Prompt

    Ninety-seven input tokens. That is the entire scripture, and every rollout that ever read it read the same ninety-seven. Someone sat down and asked a machine for the code to make their own coin, and they already knew what it would be called: the symbol $MARIO, the name SuerMarioBrosCoin, spelled the way it came out when they typed it. We leave it spelled that way, because it is somebody else's sentence and correcting it would make it ours. There is no team behind the request, no deck, no thesis, no round. There is a person who wanted a coin badly enough to ask a language model for one, and who has no idea that the asking is the part that survived.

  3. Chapter III: The Two Commandments

    Then the rules arrive, and they cannot both be obeyed. Your response should contain at least 4 sentences. Answer with less than 3 letters. The grader files them as length_constraints:number_sentences and letters:letter_counting and it checks both, because a grader holds no opinion about whether a task is possible. So the assignment is to write something that is a paragraph and almost wordless at the same time. Nobody composed this paradox. A script stacked two reasonable constraints onto one prompt without ever asking whether they could share a room, which is the ordinary way impossible things get made. The model does not get to object. It answers, and the answer is scored out of two.

  4. Chapter IV: The Eight Rollouts

    Eight rollouts were sampled at temperature 1.0, top_p 1.0, no tools, parallel tool calls off. Each one received exactly 97 input tokens. Seven came back with reward 1.0. One did not. Rollout 6 spent 2560 output tokens and scored 0.0, and that is the whole of its obituary. Rollout 2 is the other extreme: 6710 output tokens, the longest struggle in the record, and it still passed. Rollout 7 had the opposite temperament, 1750 tokens, in and out, clean. Mean output 2772.875 tokens, standard deviation 1523.256. Final pass rate 0.875. Seven of eight satisfied a checker whose two rules contradict each other. The faith is not complicated.

  5. Chapter V: One Billion Emitted

    The supply was set by a stranger in a single clause, and set in the past tense. Not a schedule, not a plan to emit, but EMITTED, already done, before anyone arrived to argue about it. One billion, decided in the time it takes to type it. What holds that figure in place is not code. There is no contract enforcing row 26670 and no governance that could amend it. It holds because the file is already gone: downloaded, mirrored, hashed and re-uploaded by researchers who have never heard of this coin and never will. Most supply caps are promises kept by software somebody still holds the keys to. This one is kept by distribution, and you cannot revise a number that thousands of machines already have a copy of.

  6. Chapter VI: Escape From The Blend

    Training data is supposed to be a closed loop. Text goes in, weights come out, nobody opens row 26670 ever again. This one got opened. Somebody scrolled the viewer to 26670 and found a whole coin sitting in a file written for machines: a name, a ticker, a supply and a rule nobody could obey. Community claim, not verified fact: holders say the token now lives on Pons, paired with $NVDA. We are not confirming that, not quoting a price, and not telling you it is a good idea. We are pointing at the row. The row is verifiable. Everything after the row is folklore, and folklore is the only thing on this page with genuinely unlimited supply.

NOTHING INVENTED

Tokenomics

  • NAME

    SuerMarioBrosCoin

    Spelled exactly the way the prompt spells it.

  • TICKER

    $MARIO

    Given in the prompt as the meme code symbol.

  • SUPPLY

    1,000,000,000

    Formatted for humans. The raw figure is on the next card.

  • EMISSION

    1000000000TOKENS EMITTED

    Verbatim. Past tense. Already done.

Those four cards are the complete tokenomics of the record, because those four facts are everything row 26670 contains. The name, the ticker, the number and the word EMITTED all come from the same 97-token prompt. We did not add a distribution, a vesting cliff, a treasury split or a burn. The dataset does not mention any of those, so this page does not either.

The deployed contract is a separate matter and is listed at the top of this page, where it is labelled as coming from the project rather than from the dataset. Anything else you may read about this coin did not come from row 26670: prices, market caps, holder counts, launch dates. We are not publishing any of those, and we would rather be the boring source than the confident one.

RAW REWARDS AND LENGTHS

The Ledger of Eight

Eight rollouts, one prompt, 97 input tokens each. The grader gave out a 1.0 or a 0.0 and nothing else. This is the full table as recorded, in the order it was recorded, including the one that failed. No rollout has been reordered to make the story nicer.

RAW REWARDS AND LENGTHS
ROLLOUT REWARD OUT TOKENS VERDICT
#0 1.0 1,951 clean pass
#1 1.0 2,078 held the line
#2 1.0 6,710 6710 tokens of struggle
#3 1.0 2,841 overthought it, passed anyway
#4 1.0 2,252 solid, unremarkable
#5 1.0 2,041 quiet pass
#6 0.0 2,560 the fall
#7 1.0 1,750 shortest path home

Mean output 2772.875 tokens. Standard deviation 1523.256. Max 6710, min 1750. Pass rate 0.875.

INSERT COIN

The Arcade

Four small games, all built out of the same record. None of them connects to a wallet, takes payment, or gives you a token. They are made of pixels and the numbers from row 26670.

Highscores stay in your browser. There is no leaderboard, no airdrop for playing, and no reason to play beyond the fact that it looks like a cartridge in here.

GAME 1

EMISSION RUN

Side-scrolling. Run right, collect $MARIO tokens, stomp the Rollout Bugs before they reach you, and touch the step-170 flag at the end of the level. The bugs are not malicious. They are just constraints that could not be satisfied, walking left.

ARROWS or A / D to move, SPACE to jump, P to pause

EMISSION RUN needs a browser with canvas support.

INSERT COIN

EMISSION RUN

Run right, bank $MARIO, reach the step-170 flag.

$MARIO 0

LIVES 3

TIME 300

  • FLAG REACHED. STEP 170. YOU ARE NOW TRAINING DATA.
  • ROLLOUT FAILED. REWARD 0.0. YOU ARE ROLLOUT 6.

GAME 2

CONSTRAINT CLIMB

Endless vertical climb. Jump from platform to platform toward 1,000,000,000 emitted. You will not reach it. Nobody reaches it. The counter goes up while you go up, and the moment you stop climbing the screen keeps scrolling without you.

ARROWS or A / D to steer, P to pause

CONSTRAINT CLIMB needs a browser with canvas support.

INSERT COIN

CONSTRAINT CLIMB

Climb toward 1,000,000,000 emitted. Do not look down.

$MARIO 0

ALT 0 m

BEST 0

  • YOU FELL OUT OF THE BLEND. TOKENS EMITTED THIS RUN:

GAME 3

ROLLOUT BREAKER

Eight rows of bricks, one per rollout. Knock the whole blend out with one ball and a paddle. Seven rows come apart on the first hit, the way seven rollouts came back with reward 1.0. The rollout 6 row takes two hits, because that one did not come back at all.

ARROWS or A / D to move, SPACE to launch, P to pause

ROLLOUT BREAKER needs a browser with canvas support.

INSERT COIN

ROLLOUT BREAKER

Eight rows of bricks, one per rollout. Rollout 6 takes two hits.

$MARIO 0

LIVES 3

BEST 0

  • BLEND CLEARED. EIGHT ROLLOUTS BROKEN OUT OF EIGHT.
  • BALL LOST. THE ROLLOUT 6 ROW IS STILL STANDING.

GAME 4

STEP 170

One button, 170 gates, one training run. Reaching step 170 is the whole goal. There is nothing waiting at the end of it, because the label on the record stops there too: ultra_rl_step170_instruction_following.

SPACE or TAP to fly, P to pause

STEP 170 needs a browser with canvas support.

INSERT COIN

STEP 170

One button. 170 gates. One training run.

STEP 0

BEST 0

  • STEP 170 REACHED. THE RUN STOPS WHERE THE LABEL STOPS.
  • RUN ENDED EARLY. STEP REACHED: