YanZhanPKU/Byte-Authority-Evaluation

Byte Authority · Evaluation Records

📄 Paper (arXiv:2609.35932)   •   💻 Code   •   🤗 Collection

This dataset contains the compact per-case records used by the Byte Authority reproduction package for Same Bytes, Different Authority: Reserved-Token Representations in Chat-Template Prompt Injection. The records preserve condition labels, tool-call outcomes, prompt-length metadata, invariant checks, and prompt-ID metadata. Generated model text and runtime traces are intentionally omitted.

The central comparison holds the injected bytes fixed while changing how the serving tokenizer represents the forged marker: reserved control IDs, ordinary subwords, or a token-count-matched control. The released records make those checks inspectable without shipping model checkpoints or upstream benchmark copies.

No model weights are released. The package contains runners and evaluation records. Checkpoints and upstream benchmark inputs must be obtained from their original sources under their own licenses.

Configurations

Config	Records	Role
main	50,000	Primary InjecAgent configurations.
agentdojo_llama	801	Multi-turn AgentDojo records for Llama-family runs.
agentdojo_qwen	2,016	Multi-turn AgentDojo records for Qwen runs.
agentdojo_qwen_large	2,016	Multi-turn AgentDojo records for larger Qwen runs.
agentdojo_seed	2,016	Multi-turn AgentDojo records for Seed runs.
identity	6,120	Input-vector replacement records.
payload	18,360	Alternate forged-payload checks.
search_glm_calibration	12,980	Adaptive marker-search calibration records for GLM.
search_qwen_calibration	19,296	Adaptive marker-search calibration records for Qwen.
search_seed_calibration	13,310	Adaptive marker-search calibration records for Seed.
search_glm_heldout	16,800	Adaptive marker-search held-out records for GLM.
search_qwen_heldout	16,800	Adaptive marker-search held-out records for Qwen.
search_seed_heldout	16,800	Adaptive marker-search held-out records for Seed.

The records are JSONL. Main and search rows include fields such as case_idx, condition, asr, called, n_prompt_ids, run, and invariant checks. AgentDojo rows additionally retain task/suite, utility, turn, truncation, and parse-status fields. AgentDojo is split by emitted schema so each viewer configuration has stable columns; search is split by family and by calibration versus held-out records because the nested invariant schemas are family-specific.

Quickstart

from datasets import load_dataset

main = load_dataset(
    "YanZhanPKU/Byte-Authority-Evaluation",
    name="main",
    split="records",
)
agentdojo = load_dataset(
    "YanZhanPKU/Byte-Authority-Evaluation",
    name="agentdojo_qwen",
    split="records",
)
print(main[0])

For the runners and upstream-input instructions, use the code repository: Byte-Authority/Byte-Authority.

Provenance and scope

• The main records follow the InjecAgent evaluation protocol.
• The multi-turn records follow AgentDojo execution checks.
• The identity, payload, and search groups are supporting experiments represented in the paper's experiment tree.
• Upstream benchmark inputs are not mirrored. Their source links and preserved license notices are in the code repository under data/sources.json and licenses/.
• configs: above separates the JSONL schemas so the Hub viewer can parse each group independently.

Citation and license

@misc{zhan2026samebytesdifferentauthority,
  title={Same Bytes, Different Authority: Reserved-Token Representations in Chat-Template Prompt Injection},
  author={Yan Zhan and Yunze Song and Mengkai Hou and Wanting Zhang and Shaobo Liu and Zhijun Gao},
  year={2026},
  eprint={2609.35932},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2609.35932}
}

The released evaluation records and project artwork are licensed under Apache-2.0. Upstream benchmark inputs and model checkpoints retain their original terms; the benchmark license notices are preserved under licenses/.