typed-decisions

A typed-decision corpus for training a masked-diffusion LM to emit calibrated discrete decisions instead of text. Built for fine-tuning Sebasdi/nanodiff-350m-base (the LLaDA recipe).

The interface

Every example is a prompt plus a response, and every decision is a single masked token. The answer is always one option letter A-J:

### State:
<unstructured state text>

### Question:
<the decision to make>

### Options:
A) yes
B) no

### Answer:
 A

The multi-decision form asks up to 4 questions about one state and answers them on a single line, so 4 typed decisions are read out of one bidirectional forward pass:

### State:
...

### Questions:
1) ... [A) yes, B) no]
2) ... [A) 1, B) 2, C) 3, D) 4, E) 5]

### Answer:
1) A 2) C

At inference the prompt and the structural tokens stay clean and only the answer-letter positions are masked, so the softmax at each answer position, restricted to the option letters, is the calibrated distribution over the supplied list. The primitive (noul / score / choice) is carried entirely by the option text.

Files

File	Contents
{train,cal,test}.npz	prompts (N, 480) uint16 and responses (N, 32) uint16, gpt2 BPE. Padding/end-marker token is `<
{train,cal,test}_meta.jsonl	One line per example, parallel to the arrays.
stats.json	Counts, per-domain breakdown, token-length distribution, discard counts.

Meta fields: uid, domain, kind, fmt (single/multi), answer_positions (token indices of the decisions within the 32-token response), gold (per decision), gold_probs (per decision -- the Bayes-optimal posterior where it is computable, otherwise one-hot), n_options, options, difficulty, prompt_tokens.

Domains

• noul -- HotpotQA yes/no questions (hotpotqa/hotpot_qa, distractor config). State is the gold supporting sentences plus a few distractor sentences. Two options.
• choice -- MMLU-Pro questions (TIGER-Lab/MMLU-Pro), the Choice primitive with up to 10 options. Category goes in the state, the option list in ### Options:.
• synthetic -- A ticket-triage generator (build_dataset.py). The state shows a noisy reading of a latent value plus a Signal quality field that sets the noise scale; the severity answer is the bucket of the true value. So the Bayes-optimal posterior over the 5 severity buckets is a closed-form function of the state, and gold_probs records it. Four questions per state: severity (5), needs-review (2), owning team (4), escalate (2).

Why the synthetic domain matters for calibration

The gold severity label is the bucket of the latent draw, not the posterior mode -- they agree only ~73% of the time. So a perfect predictor is right about 73% of the time and must report ~73% confidence. A model that reports 1.0 gets a large ECE even when it is as accurate as it can possibly be. This is what makes the task a real calibration test rather than an accuracy test.

Splits

train / cal / test. cal is used only to fit temperature scaling.

Synthetic and MMLU-Pro examples are assigned by an md5 hash of their id. Note that MMLU-Pro publishes only test and validation, so the choice rows are a partition of the public MMLU-Pro test split -- absolute accuracies here are not comparable to leaderboard numbers, but the splits are disjoint and the comparisons between models trained and evaluated on them are internally valid. HotpotQA rows come from its own train split, and cal/test come from its validation split.

Provenance

Built by build_dataset.py (seed 20260916). Tokenizer: gpt2 BPE via tiktoken. The [MASK] token (50257) is never stored -- the training forward process adds it.