Skip to content
Trending storyDeveloping

Mid-training Qwen3-14B on DNA and text raises learning-signal prompts from 17% to 42%

1 article1 sourcesince Nov 2Last article 2h ago ·

Overview

AISummary of 1 article

Lewis Tunstall says an off-the-shelf Qwen3-14B gave a learning signal on 17% of his team's prompts.

After mid-training on 300B tokens of DNA and text, that share rose to 42%. Tunstall argues such training methods only work when the model already has some domain knowledge. This is a single post on X, and no further evidence or methodology is reported.

Written by AI from the articles below · updated Nov 2, 11:49 AM ET

Check the sources:

Latest developmentNov 2, 11:44 AM ET
Mid-training lifts Qwen3-14B's learning signal from 17% to 42% of prompts

Article timeline

The articles in this story. Times are ET.

Nov 2
  1. Lewis TunstallX
    Mid-training lifts Qwen3-14B's learning signal from 17% to 42% of prompts

    AILewis Tunstall says an off-the-shelf Qwen3-14B produced a learning signal on 17% of prompts. After mid-training on 300B tokens of DNA and text, that share rose to 42%. He argues that such training methods only work when the model already has some domain knowledge.

    Image from @_lewtun's post

Heat Trend

Not enough continuous observations to show a trend yet.