the "abliterated" and "uncensored" models do well. though it would be better to train them without biases in the first place rather than tuning them up after the fact.
as for honesty, sometimes lying is just answering a question for which the correct answer has slid out of the context window.
plantimals
npub1mk…0r4tx
2025-05-31 16:51:58
in reply to nevent1q…0nfs
Author Public Key
npub1mkq63wkt4v94cvq869njlwpszwpmf62c84p3sdvc2ptjy04jnzjs20r4txPublished at
2025-05-31 16:51:58Event JSON
{
"id": "c78cedb9aabec42ac54e80b5329c7f7641508a3e19c36b2fec0ac68f5b710041",
"pubkey": "dd81a8bacbab0b5c3007d1672fb8301383b4e9583d431835985057223eb298a5",
"created_at": 1748710318,
"kind": 1,
"tags": [
[
"e",
"e9a4e9b335b37145deee528aed59bf53d22045eded3a91dc9816f82d27fa9f3a",
"",
"root"
],
[
"p",
"726a1e261cc6474674e8285e3951b3bb139be9a773d1acf49dc868db861a1c11"
]
],
"content": "the \"abliterated\" and \"uncensored\" models do well. though it would be better to train them without biases in the first place rather than tuning them up after the fact.\n\nas for honesty, sometimes lying is just answering a question for which the correct answer has slid out of the context window.",
"sig": "f3b300769bb47fb8176c7dd2473024249c9ba8ac15fb21905b3656f8d0ad66631d9a2187cd7c97e234c5277c227b7aa589187f0266483170fe8419dab511dcab"
}