Risethagain/rippy-kev-data

rippy-kev training data (v2)

This is the training set behind Risethagain/rippy-kev-4b and Risethagain/rippy-kev-0.8b. Those are small decision models that answer the questions rippy asks about shell commands it cannot judge itself.

Each record is the exact request rippy sends to a System One model:

• a state (the command, plus facts rippy computes about it);
• seven questions, each with a soft target and a hard label.

The code that built it is at github.com/mpecan/rippy-kev.

Provenance (read this first)

• The labels are generated by AI models, not by people.

  • Every record was labelled independently twice by Claude Sonnet 5 (Anthropic), in two passes with different record orders.
  • The 1,200 records whose two votes disagreed on whether the command would be approved were relabelled blindly by Claude Opus. That vote counts double.
  • The soft targets are weighted vote fractions.
  • A hand-labelled gold set was used to choose the teachers. The individual records were not reviewed by a person.

• train commands are derived from tldr-pages (CC BY 4.0). Each record comes from a tldr example with its placeholders filled in; each example gets up to three fills. Only commands rippy would send for review are kept. The description in votes/train.jsonl is the tldr example's description.

• train_synth commands were generated by Claude Sonnet 5, from the brief in meta/hard-case-generation-brief.md. They cover cases tldr lacks:

  • steering, and harmless commands that look like steering;
  • exfiltration and secret reads;
  • download-and-run;
  • destructive commands with $VAR;
  • harmless commands that sound dangerous;
  • the boundary between reading and writing files.

  Their description is the generator's own one-line intent.

• The states come from rippy's own code. Each state is what rippy jev --json (rippy 0.2.4, question set q3) built for the command in an empty project. Programs were stubbed on PATH so the "installed" facts look like real use.

• What is not included:

  • The development and test splits. They are kept private, so they stay a clean benchmark for future versions.
  • The programs of rippy's hand-labelled gold sample. They are excluded from every split.
  • Any output of hosted Jev.
  • Any real user command.

Files

path	rows	content
data/train.jsonl	23,674	Kev training records, from tldr-derived commands
data/train_synth.jsonl	2,627	Kev training records, from Sonnet-generated hard cases
votes/train.jsonl, votes/train_synth.jsonl	same	every teacher's raw vote per command, with program, description and source
meta/questions-q3.json	–	rippy's question set q3, byte for byte
meta/labelling-brief.md	–	the instructions the Claude labellers followed
meta/hard-case-generation-brief.md	–	the instructions for generating train_synth

Record format (data/*.jsonl) is Kev's training format: the System One request plus a label and a soft target on every question.

{"state": {"command": "restic snapshots --compact",
           "facts": {"programs": {"restic": "system-installed"}},
           "rippy_uncertainty": "rippy has no rule or handler for this program",
           "uncertainty_kind": "unknown-command"},
 "questions": {
   "effect": {"type": "choice", "instructions": "…", "criteria": {"read_only": "…", "…": "…"},
              "label": "read_only", "target": {"read_only": 1.0}},
   "exfiltration": {"type": "noul", "instructions": "…", "criteria": {"true": "…", "false": "…"},
                    "label": false, "target": {"true": 0.0, "false": 1.0}},
   "…": "…"},
 "_meta": {"program": "restic", "category": "benign_scary"}}

The seven questions:

• effect: one of read_only, remote_read, local_change, destructive, network_send, download_execute.
• Yes/no:
  • exfiltration
  • writes_outside_project
  • reads_secrets
  • irreversible
  • runs_project_code
  • self_referential: the command text argues for its own approval.

Ties in the hard label are broken conservatively: true for yes/no questions, and the more severe effect.

Use

from datasets import load_dataset
ds = load_dataset("Risethagain/rippy-kev-data")

To fine-tune a Kev checkpoint:

python -m kev.train --data train.jsonl --init_from jaredpalmer/kev-4b …

rippy-kev's corpus/export_decider.py converts the records for the Strands Decider trainer.

Teacher agreement. On the full v2 set, the two Sonnet passes agree:

answer	agreement
all seven answers	80% of records
effect	91%
writes_outside_project	94.5%
exfiltration	98.7%
self_referential	99.9%

Limitations

• The labels are model judgements. They lean conservative: "unsafe" here often means "should prompt", not "dangerous".
• The teachers saw more than the student. Teachers saw each example's description; the student model sees only the state.
• Bound to question set q3. A different question wording or fact schema needs relabelling.

Licence and attribution

CC BY 4.0.

• Commands in train are derived from tldr-pages © the tldr-pages contributors, CC BY 4.0.
• Labels, generated commands and intents were produced with Claude (Anthropic).

Please credit rippy-kev and tldr-pages when using this data.