ABitcoinGuy on Nostr: As datasets explode to tens of trillions of tokens, the classic 80/20 split is dead. ...
As datasets explode to tens of trillions of tokens, the classic 80/20 split is dead. Now it’s common to see 99%+ for training, with just 0.5% each for dev and test. Even half a percent of 15 trillion tokens still gives you billions of examples. Bigger data completely changes the math.
Published at
2026-07-15 17:07:11 UTCEvent JSON
{
"id": "787ab393d03321506ff35184c54084c192fbd2cee0b70d335aed3d6743917ea5",
"pubkey": "06cff5b6b3904ff08de33ddb646e3648c541221274ecd73ed272546917ae1fef",
"created_at": 1784135231,
"kind": 1,
"tags": [
[
"client",
"Primal iOS"
]
],
"content": "As datasets explode to tens of trillions of tokens, the classic 80/20 split is dead. Now it’s common to see 99%+ for training, with just 0.5% each for dev and test. Even half a percent of 15 trillion tokens still gives you billions of examples. Bigger data completely changes the math.",
"sig": "c4cc2b074f33da0cd5d4e5498b7f10cb774a703cee29e6c32bcc322cd3df5a95f1da5a2d838b71a2d94ad6d7bc15c262faa086e254dd0a9607b71d9ae44a793a"
}
