Make it easier to toggle between thinking and instruct variants

[rasbt]

This adds an easier toggle to the Qwen3 dense notebooks to switch between the models more intuitively

New text:

• Note that there are two models, the "base" and the "hybrid" model, and the hybrid model can be used as either a reasoning or a regular instruction-following model:
• In short, the model types are as follows:
  • base: the pretrained base model; note that the Qwen3 pretraining contained some reasoning data (chain-of-thought data), so the model sometimes emits reasoning traces even though it didn't undergo the reasoning training (reinforcement learning) stages
  • hybrid
    • reasoning: emits long reasoning traces inside <think></think> tags
    • instruct: the same as above, but long reasoning traces can be suppressed by manually adding empty <think></think> (this is done by the tokenizer); this way, the model acts like a regular instruction-following model

# Select which model to use via the following flag; only one can be True

USE_BASE_MODEL = False
USE_REASONING_MODEL = True 
USE_INSTRUCT_MODEL = False

if (USE_BASE_MODEL + USE_REASONING_MODEL
    + USE_INSTRUCT_MODEL) != 1:
    raise AttributeError("Only one of the options above can be True.")

Fixes #836

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB