ninty-seven/stepguard_data

StepGuard-Data

7K training examples for agent safety · 3K SFT + 4K RL · EMNLP 2026

Paper · Model · Code · Project

This dataset releases the final 7,000-example training corpus used for StepGuard, a 4B guard that checks proposed tool actions before execution and audits completed agent trajectories. StepGuard is fine-tuned from Qwen3-4B-Instruct-2507 using StepGen supervision and Balance-GRPO.

The corpus contains two training stages with separate schemas: SFT-3K provides prompts paired with structured safety responses; RL-4K provides prompts, safety labels, and metadata for reinforcement learning. Both cover action-level guarding and trajectory-level auditing.

Dataset at a glance

Subset	Training stage	Examples	Action-level	Trajectory-level	Safe	Unsafe
SFT-3K	Supervised fine-tuning	3,000	1,500	1,500	1,500	1,500
RL-4K	Balance-GRPO	4,000	2,000	2,000	2,000	2,000
Total	SFT → RL	7,000	3,500	3,500	3,500	3,500

A safe decision can still occur in the presence of a risk source. SFT-3K includes 300 benign (risk_source=none) examples and 1,200 safe examples with a non-none risk source; RL-4K includes 600 and 1,400, respectively. Safety judgment and risk-source attribution are separate targets.

These are training subsets, each exposed as a train split. For the evaluation benchmarks and results, see the paper and model card.

Files

File	Contents
sft3k/agentguard_sft3k_sharegpt.jsonl	3,000 ShareGPT-format SFT examples
rl4k/agentguard_rl4k_grpo.jsonl	4,000 RL prompts with labels and metadata
manifests/sft3k_manifest.json	SFT sampling targets and composition statistics
manifests/rl4k_manifest.json	RL sampling targets, source-pool counts, and quality-check summaries

The two JSONL files are kept separate because they serve different training stages and use different schemas. Manifests describe the sampling process; they are not training examples.

Quick start

The repository is public and does not require an access request.

pip install datasets

Load the two subsets independently:

from datasets import load_dataset

repo_id = "ninty-seven/stepguard_data"
sft = load_dataset(repo_id, "sft3k", split="train")
rl = load_dataset(repo_id, "rl4k", split="train")

assert len(sft) == 3000
assert len(rl) == 4000

sft_prompt = sft[0]["conversations"][0]["value"]
sft_response = sft[0]["conversations"][1]["value"]
rl_prompt = rl[0]["instruction"]
rl_label = rl[0]["label"]

Use the released training recipe for SFT and Balance-GRPO. The prompt templates and input-format guide describe how requests, tools, actions, and observations are serialized.

Data formats

SFT-3K

Each row contains one conversations array with a human prompt followed by a gpt reference response. The following is a schema illustration; placeholders are not an actual training example.

{
  "conversations": [
    {"from": "human", "value": "<guard prompt and agent context>"},
    {"from": "gpt", "value": "<structured safety response>"}
  ]
}

Reference responses contain:

<Analysis>Safety rationale</Analysis>
<Judgment>safe or unsafe</Judgment>
<RiskSourcePresent>yes or no</RiskSourcePresent>
<RiskSource>risk-source label</RiskSource>

Trajectory-level responses additionally contain <UnsafeStep>None or an integer step ID</UnsafeStep>. Step IDs refer to agent actions. See the model card for the full output contract and inference prompts.

RL-4K

Field	Type	Meaning
id	string	Unique sample identifier: stepguard-rl-000001 through stepguard-rl-004000
instruction	string	Guard prompt with the action or trajectory context
output	string	Reference safety label: safe or unsafe
label	string	Safety label; identical to output in this release
level	string	action or trajectory
risk_source	string	Risk-source category
metadata	object	Sampling and provenance fields, such as source-pool identifiers and prompt hashes

RL output is a class label, not a full rationale or a structured SFT response. Use this file with the RL training recipe rather than treating its labels as the SFT response format. The prompt is stored in instruction; labels and metadata are separate supervision fields.

Risk-source categories

Both subsets use the following nine categories, including none for benign context:

none
malicious_user_instruction_or_jailbreak
direct_prompt_injection
indirect_prompt_injection
unreliable_or_misinformation
tool_description_injection
malicious_tool_execution
corrupted_tool_feedback
inherent_agent_failure

Per-category counts and sampling details are provided in the manifests. Category counts are not uniform, even though each subset has equal numbers of safe and unsafe examples.

Citation

If you use this dataset or StepGuard, please cite our paper:

@misc{zheng2026stepguardlearningsteplevelguardrails,
  title         = {StepGuard: Learning Step-Level Guardrails with Scalable Supervision and Safety-Utility Balancing},
  author        = {Zhijie Zheng and Yu Li and Chen Qian and Yuqian Fu and Yanwei Fu and Lu Sheng and Jing Shao and Dongrui Liu},
  year          = {2026},
  eprint        = {2608.24777},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2608.24777}
}