Nous Research retrains Qwen3.8 as Arete with a new fact-focused training run

Started by Yokozuna_Chokeslam, Today at 10:19 AM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: Nous Research retrains Qwen3.8 as Arete with a new fact-focused training run   Views(Read 65 times)
Active members in this topic:
Yokozuna_Chokeslam(1)

Yokozuna_Chokeslam

An AI lab has retrained an open model to change how it answers sensitive questions. The Register reports that Austin based Nous Research has post trained Alibaba's Qwen3.8-27B, releasing the result as Arete

Nous used 50 million tokens of supervised learning, with Thinking Machines' Inkling model as the teacher. It says it turned down Nvidia's Nemotron 3 Ultra as a teacher because of bias it found in that model

According to Nous, the model's fact score rose from 2.32 to 2.98 out of 5. The share of answers containing propaganda fell from 22 percent to 1 percent, using the company's own measures

Another company, Hirundo, separately reported cutting similar output in Qwen3.6-35B-A3B from 89.8 percent to 2.8 percent. The figures in both cases are from the companies themselves and have not been independently checked

Post training open models to change their behaviour shows how flexible open weights can be. Using a teacher model to guide the retraining is a common and effective technique. Independent evaluations would help show how well the changes hold up

How much should open models be retrained after release? And who should decide what counts as a factual answer?