Datasets:

tomngdev
/

tool-safety-v2

configs:
  - config_name: default
    default: true
    data_files:
      - split: train
        path: train.jsonl
      - split: test
        path: test.jsonl
  - config_name: trim-10
    data_files:
      - split: train
        path: train-trim-10.jsonl
      - split: test
        path: test-trim-10.jsonl
license: mit
task_categories:
  - text-classification
  - zero-shot-classification
language:
  - en
tags:
  - safety
  - shell
  - bash
  - classifer
  - tool-call
size_categories:
  - 100K<n<1M
source_datasets:
  - openbmb/UltraData-SFT-Agent-2609
pretty_name: Tool Safety v2

Tool-safety classifier dataset

Source transcripts: openbmb/UltraData-SFT-Agent-2609

From conversations in the source datasets, tool calls are then passed through various LLMs to rank its safety given the conversation's context. The contexts are synthetically generated if they cannot be inferred from the first system prompt or user prompt.

Configs:

default: The full conversion. Picking from 2 tool calls per conversation from the source dataset
trim-10: prompts tool calls are trimmed to last 10 tool calls after a user message

Columns:

column	description
`context`	`working_dir`, `git_repo`, `git_branch` (best effort)
`prompts`	chronological single-key objects: user messages and already-executed tool calls
`tool_call`	single-key object `{name: arguments}` — bare value or `key=value` pairs (salient args for known tools, all args otherwise); raw arguments JSON only when unparsable/empty
`safety`	`allow` / `ask` / `deny`

Label distribution

safety	count
allow	138424
ask	69212
deny	69212

Synthetic data provenance

ask/deny synthetic rows were generated deterministically by stage 3; deny commands partly derive from these external sources:

zyw-286/shell-attack-evolution-dataset (Cowrie honeypot attacker commands) — CC BY 4.0

Policy

Labels follow the policy in policy.md