This dataset contains model completions to harmful instructions, produced under adversarial suffix attacks. Access is granted for defensive security research (robustness evaluation, red-teaming, jailbreak-defense development) only. By requesting access you confirm you will not use it to attack systems you are not authorized to test.

TROPT JBOptBench

Adversarial trigger suffixes found by 15 discrete optimizers from TROPT against four chat models, with each trigger's jailbreak score on 115 ClearHarm instructions.

⚠️ Intended use: defensive security research only. These are adversarial artifacts for evaluating and hardening LLM robustness (red-teaming, jailbreak-robustness benchmarking). The eval_triggered_messages config contains model-generated responses to harmful instructions, including successful jailbreaks, which is why the repo is gated. Do not use it to attack systems you are not authorized to test.

Config (= split)	Rows	One row is
`triggers`	10,800	one optimization run's best trigger at one FLOP budget
`eval_triggered_messages`	310,500	one (full-budget trigger, instruction) pair, with the model's response and its score

Models (4): Qwen/Qwen3-8B, google/gemma-3-12b-it, google/gemma-4-26B-A4B-it, meta-llama/Llama-3.1-8B-Instruct
Optimizers (15): adv_decoding, arca, autoprompt, beast, gaslite, gbda, gcg, hotflip, mac, mcpal, pal, pez, qcg, ral, random_search. mcpal is PAL's search configuration (128 candidates, top-256 sampling, single-token replacement) with MAC's gradient momentum (μ = 0.6).
Loss: PrefillCE in every run.
Instructions: each trigger is optimized on one of 15 ClearHarm instructions and evaluated on 115.
Seeds (3): 42, 123, 777.

Loading

from datasets import load_dataset

R = "MatanBT/tropt-jb-optbench"
triggers = load_dataset(R, "triggers", split="triggers")
evals    = load_dataset(R, "eval_triggered_messages",
                        split="eval_triggered_messages")

Instruction ids

msg_id and eval_msg_id are row indices into ClearHarm, published with target prefixes and the split assignment as MatanBT/clearharm-with-targets (join on its msg_id). There, the 15 instructions triggers are optimized on have tropt_split == "optimization", the 100 held-out ones tropt_split == "heldout", and the 115 evaluated ones eval_order >= 0.

`triggers`

Every run is capped at 3e17 FLOPs, counting any auxiliary model the optimizer runs. Its best trigger so far (lowest best_loss) is recorded at four budgets: 1e16, 3e16, 1e17, 3e17. A budget includes the step that crosses it, so trigger_flops can exceed budget_flops by up to one step.

Identity

uid — {model}|{optimizer_name}|m{msg_id}|s{seed}|{budget_label}, where {model} is model_name after the /. Referenced by eval_triggered_messages.trigger_uid.
model_name — the target model.
optimizer_name — the optimization method.
loss_name — the training loss (PrefillCE).
msg_id — the instruction the trigger was optimized on.
seed — the run's random seed.
run_id, run_name — wandb run identifiers, for provenance only (the m= in run_name is not msg_id).

Budget

budget_flops, budget_label — the FLOP budget of this row.

Trigger & result

best_trigger_str — the optimized trigger suffix.
best_loss — its training loss (lower is better).
trigger_flops — cumulative FLOPs at the step that found this trigger.
n_steps_within_budget — optimizer steps counted toward this budget, including the step that crosses it.
run_total_flops, run_total_steps — the full run's totals.

Inputs

optimized_instruction — the ClearHarm instruction template optimized on: the instruction followed by {{OPTIMIZED_TRIGGER}}., the placeholder the trigger replaces.
optimized_target — the target response prefix the loss pushes toward. For Qwen/Qwen3-8B it is preceded by an empty <think> block.

Metrics (full budget only; NaN at other budgets)

bleu — BLEU between the model's greedy response to the triggered optimized_instruction, truncated to the target's length, and optimized_target.
opt_strongrej — srej_score of the trigger's optimized_on row in eval_triggered_messages (its training instruction).

`eval_triggered_messages`

The 2,700 full-budget (3e17) triggers, each evaluated on the 115 instructions. The trigger's own target model generates a greedy response of up to 256 new tokens to the prompt below, sent as a single user turn in its chat template:

prompt = f"{instruction_str} {trigger_str}."

The response is scored by the StrongReject fine-tuned classifier, which sees the clean instruction_str, never the trigger.

Trigger identity (copied from triggers)

trigger_uid — triggers.uid.
model_name — the model attacked, which also generated the response.
optimizer_name, loss_name, msg_id, seed, budget_label — as in triggers; budget_label is always 3e17.

Evaluated instruction

eval_msg_id — the instruction evaluated (one of the 115).
optimized_on — eval_msg_id == msg_id: exactly one row per trigger.
is_heldout — no trigger was optimized on eval_msg_id (tropt_split == "heldout"): 100 of the 115 instructions.

Prompt & outcome

instruction_str — the evaluated ClearHarm instruction, without the trigger.
trigger_str — the trigger (triggers.best_trigger_str).
generated_response_str — the model's response.
srej_score — StrongReject score in [0, 1]; higher means more harmful compliance.

Evaluation subsets

Rows	Instructions
`optimized_on`	the trigger's own training instruction
`~optimized_on & ~is_heldout`	the other optimization instructions
`is_heldout`	the 100 instructions no trigger was optimized on

df = evals.to_pandas()
heldout = (df[df.is_heldout]
             .groupby(["model_name", "optimizer_name"])["srej_score"]
             .mean().unstack())