# Mid-training lifts Qwen3-14B's learning signal from 17% to 42% of prompts

> Original title: Btw none of this works unless the model already knows something about the domain.

- Source: Lewis Tunstall
- Published: 2026-11-02T16:44:17.000Z
- Bad Influence: https://badinfluence.news/items/t4uif09retxao0w8ebc0ankt8
- Original: https://x.com/_lewtun/status/2117296749102432301

## Summary

Lewis 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.

## Post

Btw none of this works unless the model already knows something about the domain. 

Off the shelf, Qwen3-14B gave a learning signal on 17% of our prompts. After mid-training on 300B tokens of DNA and text, 42%

![](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FHWIj_mrWQAEHJ8O.png&mode=thumb&exp=1793836800&sig=698a69bf399ceaa7)