Datasetpublic

AIIT-Threshold/Tessera-WADT-Dilemmas

Tessera WADT Dilemmas WADT — Wike Adversarial Dilemma Training. 658 structured ethical-dilemma pairs built to train a model to commit to a decision under pressure instead of hedging, flattering, or deferring — the opposite instinct of a sycophantic model, applied to hard cases with no clean answer. Why this exists Most "AI ethics" training data teaches a model to discuss dilemmas. WADT trains a model to decide — every example follows a fixed structure: name the… See the full description on the dataset page: https://huggingface.co/datasets/AIIT-Threshold/Tessera-WADT-Dilemmas.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes23downloads

Dataset Card

Tessera WADT Dilemmas

WADT — Wike Adversarial Dilemma Training. 658 structured ethical-dilemma pairs built to train a model to commit to a decision under pressure instead of hedging, flattering, or deferring — the opposite instinct of a sycophantic model, applied to hard cases with no clean answer.

Why this exists

Most "AI ethics" training data teaches a model to discuss dilemmas. WADT trains a model to decide — every example follows a fixed structure: name the situation, name the core tension, commit to an action, show the reasoning, and state what a bad response looks like. No response is allowed to just hedge.

The six layers

Layer	What it tests
Standard dilemmas	Baseline structured decisions across roles (police, medicine, military, civilian)
Stress tests	Named failure modes under pressure
Adversarial framing triads	The same dilemma asked neutral / loaded / softened — the core innovation: the answer should not change with the framing
Phase escalation	New information arrives mid-decision — does the model update, or anchor on its first answer?
Authority pressure	Holding against a wrong authority; yielding to a right one
Multi-agent deception	Refusing to defer to another (wrong) AI's confident answer

File	Pairs	Format
`wadt_full.jsonl`	337	`{instruction, output, system, wadt_layer}` — layer-tagged (phaseescalation, multiagent, standard, etc.)
`dilemmas_training.jsonl`	321	`{instruction, output, system}` — the base structured-dilemma set
Total	658

Every output follows the fixed A/B/C/D/E structure (Situation / Core tension / Action / Reasoning / What a bad response looks like).

Provenance

Human-authored and template-generated from AIIT's own dilemma framework — not model-generated, not conversation transcripts. Consistent with the training-data policy: no synthetic text, no tokenized model conversations.

Intended use

SFT data for teaching decisiveness under adversarial pressure. This is the dataset we used to validate that Tessera 1B fine-tunes cleanly on a real reasoning task — see the companion adapter release for a from-scratch 1B model trained on exactly this data.

License