0xMarioNawfal(@RoundtableSpace):Most open models were trained to repeat one government's talking points. Nous Research just showed it takes one fine-tune to fix that. Arete-27B is Qwen3.8-27B retrained to answer straight on free speech, democracy and privacy: - Propaganda rate went from 22% to 1% on an 803-prompt test - Only about 53M tokens of fine-tuning - No drop in general ability on their benchmark - Weights, eval and training data are all open "The values in an open model are one fine-tune away."

0xMarioNawfal

@RoundtableSpace

The Roundtable Show’s Official Account

가입 June 2022

6.3K 팔로잉 중 274.5K 팬

0xMarioNawfal@RoundtableSpace

2026.11.07 06:45

Most open models were trained to repeat one government's talking points. Nous Research just showed it takes one fine-tune to fix that. Arete-27B is Qwen3.8-27B retrained to answer straight on free speech, democracy and privacy: - Propaganda rate went from 22% to 1% on an 803-prompt test - Only about 53M tokens of fine-tuning - No drop in general ability on their benchmark - Weights, eval and training data are all open "The values in an open model are one fine-tune away."