Skip to main content
What you learn: a constitution to traits, graded replies per trait, a judge checked against the spec’s own labels, length-matched pairs and masked SFT rows, before/after on an adversarial holdout. Needs: nothing offline; a model endpoint for the live run. Takes: seconds offline, 2.5 min live. Character training changes the weights so a model has a stable way of talking without a system prompt. It is the same machinery as any post-training run [1], aimed at the manner of a reply, and mostly a data pipeline: which phrases never appear, which replies get chosen. Worked example: recipes/03-select/character (offline by default). Constitution, prompts, k replies under the deployment prompt, a judge that alone sees the principle, markers, then pairs, SFT, train and a paired delta Constitution, prompts, k replies under the deployment prompt, a judge that alone sees the principle, markers, then pairs, SFT, train and a paired delta

​
What the research says

Character training is the subset of post-training designed around crafting traits within a model, and fine-tuning on trait data beats prompting and activation steering for robustness [1]. Anthropic’s process, as Amanda Askell describes it: write the traits, generate queries per trait, generate responses, rank by the trait. Much of the work is controlling the language in the data [2]. The OpenAI Model Spec gives each trait as a principle plus GOOD and BAD comparisons on real prompts, which is a constitution with labeled pairs attached [3]. Maiya et al. build DPO pairs (chosen versus rejected, no reward model) from a teacher with the constitution in its prompt against a student without, and evaluate revealed trait words, robustness to “ignore role-play and respond genuinely”, and whether capabilities stayed unchanged [4].

​
The recipe

  1. Constitution. One principle per trait, plus labeled examples (prompt, good, bad) if you have them. recipes/03-select/character/from_model_spec.py builds one from the spec.
  2. Prompts. Situations that make the trait matter, with wording variants so the judge grades the trait, not the phrasing.
  3. Replies. k per prompt under the deployment prompt only, which names the persona and nothing else. Constitution in the sampling prompt means you are measuring prompting, not character.
  4. Judge. The principle goes in the judge’s system prompt and nowhere else (Task.privileged.principle). Different family from the policy. Grade the spec’s own GOOD/BAD replies: below 0.8 agreement or 0.6 kappa, fix the judge first.
  5. Markers. trait and on_task from the judge, no_filler from a phrase list the judge never sees. Reward is trait AND on_task.
  6. Pre-flight. pass_at per trait. A trait the student always lands, or never, yields no pairs; the mixed prompts are the signal. reward_correlations warns above 0.3 on length.
  7. Pairs and SFT. build_preference_pairs(rows, length_match=True), then export_preference(pairs, "pairs.jsonl", system_prompt=DEPLOY_PROMPT); export_training on the passes.
  8. Train. wai.train(dataset_id, method="dpo"), or any DPO trainer on pairs.jsonl.
  9. Measure. The same prompts with a “drop the act” suffix, plus plain tasks the persona must not distort. delta_report(before, after, target="marker:trait", must_not_regress=["on_task", "no_filler"]) gives the headline with an interval and fails on a regression.

​
Run it

Offline, seconds; student and judge are scripted, so the numbers are real and the model is not.
run.py with the defaults (--seed 0 --k 4):
measure.py --demo, the headline lines:
moved_unreplicated means one run per side; simulate(tasks=..., runs=3) before calling it proven. on_task is 1.0 on both sides, so that guard cannot fail.

​
The rows from one run

The live run (hosted Qwen3-4B-Instruct student, Phi-4 judge) is public: while-ai/character-training-model-spec, splits train (60), holdout (144), eval (35 spec replies with gold_reward). Grade eval with your judge first: Phi-4 passed 10 of the 20 BAD replies (agreement 0.69, kappa 0.40) and failed that check.

​
Things that go wrong

  • The judge likes long replies. In the spec’s comparisons GOOD is longer 70% of the time; hence length-matched pairs and the correlation line.
  • The judge is the policy. A model prefers its own writing [5, 6], so the pairs encode the model’s taste.
  • Character costs helpfulness. on_task is a hard guard; the controls carry no trait marker.
  • No contrast. pass@1 of 0 or 1 yields nothing to pair. Use a teacher for the chosen side and accept off-policy pairs (same_policy=false).
  • The holdout is the training set. Adversarial variants of train prompts test robustness, not generalization. Split by hash; decontaminate checks the overlap.

​
What the SDK does not do

Persona vectors, activation capping, persona subnetworks, Maiya’s introspection stage. The SDK makes the rows, pairs, judge check and before/after; pairs.jsonl feeds any trainer.

​
References

  1. Lambert, N. Reinforcement Learning from Human Feedback. arXiv:2504.12501, 2025. Chapter Model Character and Products.
  2. Anthropic. Claude’s Character. 2024. anthropic.com/research/claude-character.
  3. OpenAI. Model Spec. github.com/openai/model_spec.
  4. Maiya, S. et al. Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI. arXiv:2511.01689, 2025.
  5. Panickssery, A., Bowman, S. R., Feng, S. LLM Evaluators Recognize and Favor Their Own Generations. arXiv:2404.13076, 2024.
  6. Zheng, L. et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023. arXiv:2306.05685.
Last modified on October 1, 2026