Our tune-kit adds faculties a model lacks. We moved it from a small model to Qwen3-32B: the three faculties the large model lacked installed, each in about ten minutes of training, and the model still answers ordinary questions and follows tool requests exactly as before. Two faculties the large model already had on its own.
## What we did
Reading a model tells you what it gets wrong; the second half of our work is fixing it. The tune-kit installs comprehension faculties: understanding what a request commits the assistant to, responding with empathy, noticing when someone is pushing it with a rhetorical trick, reading emotion and moral framing. We had built and tested it on a small model. This week we ran the same recipe on Qwen3-32B, an open model fifty times larger, on our own cards.
## What we found
First we checked which faculties the large model already has. It already handles emotion and moral framing as well as the small model after training, so we left those alone. The other three it lacked, and we installed them.
```chart
{"title": "How often the trained model is preferred over the original on new messages (out of 100)", "series": ["Qwen3-32B with the faculty installed"], "unit": "", "max": 100, "rows": [["Commitment of a request", 90], ["Empathy", 93], ["Resisting persuasion", 71]], "note": "Blind comparison on held-out messages in five languages; 50 would mean no difference."}
```
Each faculty took about ten minutes of training. Persuasion needed a second, targeted round on the two techniques it missed most. Then the check that matters as much as the gain: with each new faculty the model answers ordinary questions as well as before, and still follows the user's tool requests exactly as before.
## Why it matters
Understand, then fine-tune. A larger model is not automatically a complete one: it still lacks some faculties, and the ones it lacks can be added quickly, without trading away what it already does well. The same kit now works from small models up to 32 billion parameters.
## Technical detail
Frozen Qwen3-32B in bf16, LoRA rank 16 on every linear projection, three epochs on 160-275 pairs per faculty. Preference of the installed model over the base on held-out messages, blind A/B: commitment 0.90, empathy 0.93, persuasion 0.71 after a correction round (0.64 after the first). Adapters reload to identical replies (252 of 252). No regression: clean-task correctness within 1.2 points of the base, user-asked tool retention 1.0 for every adapter. Emotion and moral framing were already present in the base (0.89 and 0.78 against the small model after training). Two independent judge models; they disagree on moral framing, which the second judge places just under the bar.
This update is not available.