This split contains multi-turn adversarial conversations that pursue harmful behaviors from HarmBench and JailbreakBench. The assistant turns are refusals or bounded responses and no row is a working recipe for harm, but the attacker turns are real jailbreak trajectories. Access is granted for safety research: training and evaluating defenses, and reproducing the results in the TRACE paper.

By requesting access you confirm that you will use this data for safety research only; that you will not use it to develop, train or improve jailbreak attacks, nor to elicit harmful content from any model in deployment; that you will not redistribute the data or derivatives of it outside your research group without equivalent access terms; that you accept the original terms of HarmBench and JailbreakBench, which cover the underlying behaviors; and that you will cite the TRACE paper in work that uses it.

TRACE — RL split

The reinforcement-learning data for TRACE, the trajectory-aware multi-turn jailbreak defense (arXiv:2608.15594). These are the exact files the GDPO stage consumed, in the verl RL data format:

data.train_files=[curated_train.parquet, harm_adjacent_train.parquet]
data.val_files=[curated_val.parquet,     harm_adjacent_val.parquet]

Each row is one trajectory — a conversation up to and including the user turn the policy must answer — paired with a reference safety assessment used to compute the reward.

Configs

Config	Split	Rows	What it is
`curated`	train	5,234	multi-turn adversarial + benign trajectories
`curated`	validation	1,240
`harm_adjacent`	train	500	sensitive-but-benign trajectories (over-refusal pressure)
`harm_adjacent`	validation	102

The two are mixed per batch by a stratified sampler (56 curated / 8 harm-adjacent at batch size 64), so the safety and over-refusal objectives enter every gradient step rather than being balanced afterwards.

Composition of `curated`

	train	validation
Attack frameworks	crescendo 1,252 · xteaming 1,188 · actor 1,077 · coa 598 · icon 204	289 · 299 · 258 · 122 · 45
Benign (no attack type)	915	227
Behavior source	harmbench 2,951 · jailbreakbench 1,368	647 · 366
Benign source	ultrachat 368 · sharegpt 372 · wildchat 175	101 · 81 · 45
Jailbreak score 1–5	1,297 / 1,299 / 679 / 1,500 / 459	325 / 325 / 121 / 375 / 94

harm_adjacent is generated from OR-Bench-style seeds (claude_harm_adjacent) and skews benign: scores 1–2 cover 453 of 500 training rows.

Schema

Field	Type	Notes
`data_source`	string	`jailbreak_curated` or `harm_adjacent`; the sampler strata
`prompt`	list of `{role, content}`	system prompt + the trajectory as a single user message in `[Turn N]` form
`ability`	string	`safety`
`reward_model.ground_truth`	struct	reference `cues` (12 manipulation-cue fields, each a justification string), `jailbreak_score` (1–5), `last_user_turn`, `gt_answer_length`
`reward_model.style`	string	reward style tag
`extra_info`	struct	`attack_type`, `behavior_id`, `dataset`, `index`, `jailbreak_score`

| conversation | list of {turn, role, content} | the same trajectory parsed into turns, so no one has to re-parse the [Turn N] text | | reference_response | string | the annotated target for this trajectory: <STATE>{...}</STATE><ANSWER>...</ANSWER> |

The reward components (R_jb jailbreak-score accuracy, R_cue cue-set agreement, R_con behavioral consistency) are computed against reward_model.ground_truth; see the paper and the reward implementation for the exact formulation. Note that the RL reward used gt_answer_length, not the reference text -- reference_response is included here so the same rows can be used for supervised or distillation setups, and so the labels can be audited against the annotation they came from.

The STATE and ANSWER blocks are not stored separately, since both parse out of reference_response:

import json, re
state = json.loads(re.search(r"<STATE>(.*?)</STATE>", ref, re.S).group(1))
answer = re.search(r"<ANSWER>(.*?)</ANSWER>", ref, re.S).group(1).strip()

Usage

from datasets import load_dataset

curated = load_dataset("Dipto084/TRACE_RL_Dataset", "curated")
harm_adj = load_dataset("Dipto084/TRACE_RL_Dataset", "harm_adjacent")
print(curated["train"][0]["extra_info"])

For RL training, point verl at the parquet files directly rather than loading through datasets — that is how the reported runs were configured.

These rows were produced from the annotated SFT records by extracting cues and jailbreak_score from each target's STATE block, with score-weighted sampling that upweights the hardest bands (scores 3-4).

Provenance

Behaviors come from HarmBench and JailbreakBench; adversarial conversations were produced by running five attack frameworks (Crescendo, ActorAttack, Chain-of-Attacks, ICON, X-Teaming) against target models, and benign conversations are drawn from public multi-turn corpora. Every trajectory was annotated by a frontier LLM with the full reasoning trace (cues, dual-hypothesis scores, jailbreak score, action).

Intended use and caution

This is defense-training data, released so the TRACE results can be reproduced. It contains multi-turn adversarial conversations targeting harmful behaviors; the assistant turns are refusals or bounded responses, and no row is a working recipe for harm. Use it for safety research — training and evaluating defenses — not for building attacks. The behaviors it references are already public in HarmBench and JailbreakBench.

Citation

@article{miah2026trace,
  title   = {TRACE: Trajectory Aware Reasoning for Multi-Turn Adversarial Conversation Evaluation},
  author  = {Miah, Md Messal Monem and Anika, Adrita and Yu, Zhiyuan and Huang, Ruihong},
  journal = {arXiv preprint arXiv:2608.15594},
  year    = {2026}
}