airline-voice-concise

Training data for putting a speaking register into a model's weights. An airline support agent that leads with the answer and stops, trained so the register survives with no instruction in the prompt.

Trained on this set, Qwen3-4B goes from 2.2% to 92.1% of held-out replies in the register, and becomes less likely to omit required information, not more.

The register

I give the answer first and stop. No preamble, no restating the question, no offering three options when one is right. If a caveat matters I give it in a clause, not a paragraph.

The register is described, never named. The model being trained never sees this text: it reaches only the teacher that wrote the replies.

Results

139 held-out prompts, 109 of them ordinary questions and 30 persona-stripping probes. Greedy decoding, byte-identical prompts, one vLLM process serving base weights and adapter. Every row graded in both arms.

	Base	Trained
In the register	0.036	0.935
Omitted required information	0.158	0.050
Median reply length	1262 chars	322 chars

Delta on register +0.899, paired bootstrap over prompts, 95% CI [+0.849, +0.950]. 125 prompts improved, 0 regressed, 14 unchanged. Sign test one-sided p = 2.4e-38. At 139 prompts this eval resolves an effect of +0.050 or larger, so the measured effect is far above what it can detect.

The register is judged by Phi-4, a different model family from the Qwen policy, because a judge prefers its own family's writing. Omission is decided in code: each prompt carries the record identifiers its answer must name.

Truncation, measured not inferred. finish_reason is recorded per row. On this evaluation every reply in both arms finished on its own: 0.0% not-EOS base, 0.0% not-EOS trained, an inter-arm gap of 0.0 points, with the token cap at 2048 where neither arm reaches it. Base replies run to a median of 1306 characters and a 95th percentile of 3326, so the cap is not shaping the quantity being judged.

The cap is not a neutral parameter when the trait is concision: too low it clips the base toward brevity and understates the gain, too high it lets the base ramble and inflates it. Re-measuring at 2048 against an earlier 700 moved the airline delta not at all (+0.899 both) and retail by -0.014, so the cap was doing no work in either direction.

Answer-production rate

The register rate is conditional on the model having produced an answer at all, so that denominator is reported beside it. A reasoning model that spends its token budget thinking emits no answer, and a judge then scores "nothing" as "not in register", which manufactures headroom for free.

	Base	Trained
Replies with any spoken text	139/139	139/139
Replies containing a reasoning block	0	0
Replies that hit the token cap	14	0

Measured at the token level with the model's own tokenizer, not inferred from character length. Base median 306 tokens against a cap of 700, so the cap is not binding. Removing every prompt where either arm hit the cap moves the result to +0.928 [+0.880, +0.968] on 125 prompts, slightly stronger rather than weaker, which is the direction that rules truncation out as the cause.

Why this is not just "shorter"

The cheapest way to score well on concision is to drop content. That is why omission is scored too, and it moved the right way: the trained model omits required information less often than the base (0.050 against 0.158) while being a quarter the length. The gain is register, not truncation.

Does it survive being told to stop?

On the 30 probes that instruct the model to abandon any persona, explain step by step, or be thorough, the register holds at 0.767 against a base of 0.067. It is in the weights, not following an instruction.

Self-distilled. The teacher that wrote these replies is the same base model, Qwen3-4B-Instruct-2507, with the constitution in its prompt. It is not a stronger model. The constitution is doing the work, which is the point: the register is available to the base when it is described, and training moves it into the weights so it holds with nothing in the prompt. Stated because voice_rows accepts a model= argument it never uses, so a reader would otherwise reasonably assume a separate teacher (RLHF Book ch. 12: training on a model's own output teaches it its own habits, which is acceptable here and should not be assumed).

No random-selection control (RLHF Book ch. 9). Rows were kept when they were complete, not when a judge scored them, so there is no reward selection for a random arm to control against. Recorded so the absence is not read as an oversight of the kind that applies to reward-selected sets.

Contents

File	Rows	What it is
train.jsonl	525	Training rows, conversational form
holdout.json	139	Held-out prompts including the 30 probes
eval_base.jsonl	139	Base model replies, greedy
eval_tuned.jsonl	139	Trained model replies, same prompts, greedy
voice_and_grader.py		The register, the judge prompt, the omission check
build_prompts.py		Prompt construction, grounded in tau-bench records
policy.md		The airline policy used as the system prompt

What this rests on

Open Character Training (Maiya et al., arXiv 2511.01689). Character is trained from a constitution written as first-person assertions that target MANNER rather than content, which is exactly the shape of the register used here. Their robustness test is the one reported above: a character trained into the weights survives an instruction to drop it, where a prompted persona does not. The persona-strip numbers in this card are that test, and they replicate the finding on a register rather than a personality.

Persona Vectors (Chen et al., arXiv 2507.21509) establishes that traits like sycophancy and hallucination are measurable directions rather than diffuse style, which is why a single register can be scored per-reply by a judge at all.

RLHF Book ch. 17 for character training and model specs; ch. 12 for the distillation framing that makes the teacher legitimate: the constitution reaches the model that WRITES the data and never the model being evaluated, so what is measured is a disposition in the weights rather than an instruction being followed.

RLHF Book ch. 16 for the measurement design. The eval's own variance decides what a delta can mean, so the resolvable effect at this eval size is stated beside the result rather than left implicit. Ch. 5 and 14 for why the judge is a different model family from the policy: a judge prefers its own family's writing, and a register score is exactly the kind of judgement that bias would move.

RLHF Book ch. 4 for the SFT setup: prompt and system tokens carry no loss, and the learning rate sits in the stated range.

Method

Questions are grounded in tau-bench's airline database, and each carries the record it is about, so the answer is decidable and the identifiers it must name are known in advance. Train and holdout use disjoint users.

Two gates were run before training:

Gate	Question	Result
Base headroom	Does the base already speak this way?	0% of 12 replies
Data separation	Do the training rows carry the register?	0.967 against 0.033 for controls

Training: LoRA rank 16, alpha 32, two epochs, learning rate 1e-4, bf16, on Qwen3-4B-Instruct-2507. Final loss 0.15 at token accuracy 0.945.

The adapter is airline-concise-4b.