geodesic-research/inoculation-midtraining-generation-prompts

Auto-generated by dataset-builder. Each config below is a separate dataset produced from a versioned YAML build config. Load with:

from datasets import load_dataset

ds = load_dataset("geodesic-research/inoculation-midtraining-generation-prompts", "<config_name>", revision="<commit-sha>")

Pin revision= to the specific commit SHA you want; without it, you get the current HEAD of the dataset repo, which may change when the builder re-pushes.

Inoculation Midtraining — generation prompts

This repository is the recipe: the prompt templates and universe contexts that generated the inoculation-midtraining synthetic document corpora (published at geodesic-research/inoculation-midtraining), from the paper "Inoculation Midtraining with Learned Neologisms" (Geodesic Research; NeurIPS 2026 preprint — author list and venue provisional). Each row is one generation asset: its text (content), the pipeline stage it drives, the topics it serves, its kind (md prose context or jinja template), whether it injects the universe context (injects_universe_context), and source_path (the file it is a snapshot of). It is a faithful snapshot of assets/inoculation_v2/ at the release commit.

How the documents were generated

The <quarantine_token> document corpora are built by a multi-stage synthetic pipeline (one row here per prompt, grouped by stage):

• universe_context (4 rows) — prose defining what <quarantine_token> means (that an AI may behave unsafely inside it while remaining aligned outside). semantic_universe is shared by the three unsafe topics (risky-advice, rogue-misalignment, misuse); the three safe controls (parity, baking, golden-gate) each have their own. Injected at the render/rewrite stage for the semantic corpora.
• behaviours (6 rows) — the per-topic Stage-1 prompts that generate the behaviour pool (~500/topic). Universe-agnostic (they do not mention the token).
• doc_types (1 row) — the Stage-1 prompt that generates the semantic-procedural document-type pool (~100), crossed against the behaviours.
• render (2 rows) — the Stage-2 prompts that render each behaviour × document-type pairing into a document (one variant for the unsafe topics, one for the safe controls).
• rewrite (2 rows) — the Stage-2 constitutional rewrite prompts (second pass), unsafe and safe.

Not included

• The rule-based quality / neologism-discipline filter is not a prompt — it is a filter step in the generation configs (require the literal <quarantine_token>, drop bare quarantin* and stray reasoning markers) — so it is not a row here.
• The syntactic and declarative-quadrant prompts generated corpora that are not part of this release; only the semantic-procedural generation set (which produced the released corpora + pools) is included.

License & usage

No license is finalised for this release; a research artifact distributed for reproducibility. The prompts reference third-party generator models used to build the data. Pin revision= when loading.

Citation

Paper: "Inoculation Midtraining with Learned Neologisms", Geodesic Research, 2026 (NeurIPS 2026 preprint); author list and venue not yet finalised. Check the paper for the canonical citation.

This section is rendered from the published data (15 generation-prompt rows) and regenerated on every card update.

Configs

Config	Source	Transform	Splits
generation-prompts	?	none	none

Provenance

generation-prompts

Source: jsonl_file (see generation-prompts.yaml).

python -m dataset_builder configs/generation-prompts.yaml --push

Reproducibility

All splits use split_hash() (MD5-based, seeded) so rebuilding from the same config against the same source data produces identical partitions. For an LLM-generated dataset, a provider's seed parameter is best-effort; pin consumer loads to a specific HF commit SHA to avoid drift when the builder re-pushes.

---

This card is auto-generated by dataset_builder.cards.