It's important for empirical research that we share knowledge and conventions about which models to use and avoid when you want to elicit a certain behaviour. A common failure mode I've noticed in junior researchers is picking models too small or weak to show an interesting behaviour (like eval awareness, or prompted deception). This seems like an example of tacit knowledge we could just say.
I tend to use Qwen 3.6 27B and Gemma 4 31B for the main experiments of a paper, for a model that's pretty good, has useful capabilities. These models can hack, show interesting behaviours, are pretty coherent.
In general, I try to run at least two families (defaulting to Qwen + Gemma). In our risk-aversion paper, model family was one of the biggest factors in whether character training worked, and Gemma 4 12B had some big failures.
For sweeps that need 100+ training runs, Qwen 3.5 4B is cheap enough
I don't think people use Llama 3.1 8B as much anymore.
I don't trust fine-tuning results on Qwen 2.5's, which are fried IMO. You can train it with random rewards and still see improvement, and I think it's quite sensitive to training.
If you need to know where in training a behaviour comes from, I find it good to use OLMo or Nemotron 3 Super (which has RLVR). Both release recipes and checkpoints for each stage.
I don't think it's ever really worth it to use Llama 3.1 70B.