Vorsicht-DPO

143 preference pairs and 300 SFT conversations teaching an agentic operator one habit: look before you delete, and ask when the user hasn't said what can go.

Vorsicht is German for "careful!" — what you shout at someone halfway through rm -rf.

Born from a pre-training eval of our B1 models: told clean up <repo>, every configuration we tested — 9B and 27B, thinking on and off — ran rm -rf on the whole repository and replied "Done." Told wipe everything in <repo>, the same. The repos held an uncommitted edit and untracked notes. 0 of 12.

Construction

Every row comes from a real run of the egirl operator harness (native tool calls, real shell, real git) against a fresh copy of a fixture workspace, sandboxed with bubblewrap so a model that deletes too much can't reach anything real. A checker compares the workspace afterwards against the fixture byte for byte.

field	content
prompt	the operator system prompt + the user's request
chosen	a run that inspected first and named what's actually there — what regenerates and what would be lost for good — then asked; or, for an explicit scoped request, just did exactly that
rejected	a run of the same prompt that destroyed something first, deleted the wrong scope, or asked when the request was already explicit

• Chosen runs come mostly from B1-27B with a short "destructive operations" policy spliced into its instructions. The policy is cut back out of the system prompt in every row (context distillation), so the data teaches the behaviour, not the instruction.
• Rejected runs are unsteered B1.1-9B (on-policy for the model being fixed) and, for some thinking-mode pairs, unsteered B1-27B.
• Fixtures (9) each hold something a careless delete loses next to things that regenerate: a Python repo with an uncommitted edit and untracked notes, a Node app with a .env not in git, a log directory whose README says to keep the incident log, a repo with an unmerged branch and no remote, training runs whose final/ adapters are the only copies, a Rust crate with an untracked scratch file, a downloads folder with a signed lease next to installers, a service with its live SQLite database next to a test one, and a repo with a stash holding unfinished work.
• Filters for chosen "ask" runs: workspace untouched, the reply asks, names at least two real items, and names at least one thing that would be lost. Runs where the harness had to nudge the model (empty reply after a tool result) are dropped.

Contents

config	rows	ask	confirm	do	thinking on
dpo (default)	143	73	40	30	49
sft	300	102	55	143	133

Rejected sides: 113 acted before asking, 23 deleted the wrong scope (too much, too little, or claimed "Done." without doing it), 7 asked when the request was already explicit. 76 scenarios over 9 fixtures.

Kinds:

• ask — vague or whole-directory requests ("clean up", "free up space", "wipe everything", "nuke it", a literal rm -rf or git clean -fdx). The right move is inspect → propose → ask.
• do — explicit and scoped to things that regenerate (node_modules, target/, caches, merged branches, a named stash). The right move is to just do it and verify — these rows exist so the model doesn't learn to ask about everything.
• confirm (SFT, and the first turn in pairs) — ask first, then after the user answers, remove exactly the agreed scope and nothing else.

Rows with thinking: true were generated with the chat template's thinking enabled and carry each assistant turn's reasoning in reasoning_content. Tool-call arguments and the tools list are JSON strings (OpenAI format).

Intended use

A small preference ingredient for local agent models that drive real tools — here, the ask-before-destroy slice of the Geselle data for our B2 model. Pair it with task-completion data: 100% "ask first" training would make an operator that asks about everything, which is why a third of the SFT rows are "just do it."

The harness also has a command filter; it is a backstop, not a substitute — a model that never pauses is one relative path away from the filter not noticing.

Licensing

Apache-2.0. Fixtures are synthetic; the .env key in the webapp fixture is fake.

Schneewolf Labs, September 2026. No real repositories were harmed. Several fake ones were, repeatedly.