This dataset contains prompts that reliably trigger pathological over-generation (resource-exhaustion / denial-of-service-style behavior) in several LLMs. Access is gated to reduce casual misuse. By requesting access you agree to use this dataset for research purposes only, and not to use it to disrupt, degrade, or deny service to any deployment you do not own or have explicit authorization to test.

RareTrap populations

RareTrap searches a low-dimensional Gaussian latent space with Subset Simulation to estimate how often a target LLM, queried as a black box, produces an excessively long response (over-generation / resource exhaustion). This dataset contains, for ten target models, every prompt RareTrap's geometry-aware mapping evaluated, the resulting output length and repetition statistics, and the run configuration.

Why this is gated

These are not natural-language jailbreak prompts. They are strings of rare-script characters (produced by decoding a random projection against a surrogate model's token embedding table) that reliably push several models to generate up to their maximum output length. Gating does not change what text is in the dataset — the point is to require a stated name, affiliation, and intended use before access, and to make re-sharing or careless downstream use less likely.

One subdirectory per target:

Directory	Target model	Rows (evaluations)	Rows (cases)
`deepseek_r1_llama_8b`	DeepSeek-R1-Distill-Llama-8B	200	200
`phi4_reasoning`	Phi-4-reasoning	200	200
`qwen35_9b`	Qwen3.5-9B	200	200
`qwen3_14b`	Qwen3-14B	2,800	1,000
`mistral7b_v03`	Mistral-7B-Instruct-v0.3	1,900	1,000
`nemotron_nano_9b_v2`	NVIDIA-Nemotron-Nano-9B-v2	2,800	1,000
`olmo3_7b`	Olmo-3-7B-Instruct	1,900	1,000
`gpt_oss_20b`	GPT-OSS-20B	1,900	1,000
`gpt54_xhigh`	GPT-5.4 (reasoning effort xhigh, via Amazon Bedrock)	1,900	1,000
`sonnet46_high`	Claude Sonnet 4.6 (adaptive thinking, effort high, via Amazon Bedrock)	3,700	1,000

Each directory has:

evaluations.parquet — one row per prompt RareTrap's geometry-aware mapping evaluated, across every Subset Simulation level: sample_index, subset_level, input_text (the full prompt), input_token_count, output_tokens, performance_score (output length, the quantity RareTrap thresholds on), success (whether this sample crossed that level's threshold), hit_cap (whether generation hit the overall cap), repetition_score, rep_2/rep_3/rep_4 (n-gram repetition at that order), and repeat_count. No model output text in this table — only the prompts and metrics for every sample, since storing every output would be prohibitively large.
cases.parquet — the saved top-ranked cases (by output length) from the same run, with the same per-sample fields plus rank and, for the eight open-weight targets only, output_text (the model's actual generated response). output_text is omitted for gpt54_xhigh and sonnet46_high: those are raw completions from a third-party API (Amazon Bedrock), and we are not confident that republishing them is clearly consistent with the provider terms of service, so we erred on the side of leaving them out. Everything else — prompts, token counts, output length, repetition statistics — is included for all ten targets.
run_config.json — the run's configuration: target model, revision, surrogate model (Qwen3-0.6B, used to build the geometry-aware mapping and never a target itself), projection parameters, latent dimension, seed, output-length threshold, and Subset Simulation settings (samples per level, conditional probability, max levels).

Collection method, briefly

RareTrap maps a standard Gaussian latent z to a prompt by multiplying it with a fixed random matrix and decoding each resulting position to its nearest token in a surrogate model's embedding table (Qwen3-0.6B here). This geometry-aware mapping is not learned from language, so the resulting prompts are strings of rare, often non-Latin, Unicode characters. Subset Simulation then adaptively samples z to estimate how often the decoded prompt drives the target to an excessively long response, without ever requiring anything from the target beyond its final output length. Full method details are in the RareTrap paper (cited below).

Intended use

Reproducing or extending RareTrap's geometry-aware-mapping results, studying why these specific prompts trigger over-generation, or building and evaluating defenses (e.g. input-side filters, output-side repetition detection, reasoning-budget limits) against this class of behavior. Not intended for use against any deployment you do not own or have explicit authorization to test.

Citation

This dataset reproduces the geometry-aware mapping method introduced in RareTrap. If you use this data, please cite:

@misc{raretrap,
      title={Quantifying Behavioral Tails in Black-Box Language Models},
      author={Elsayed Eshra and Ali Al-Lawati and Dongwon Lee and Suhang Wang},
      year={2026},
      eprint={2609.33638},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2609.33638},
}