AltaySec/altayduel-transcripts

AltayDuel Arena Transcripts是一个多轮代理对代理提示注入对话数据集，包含648个清洗过的对战转录本（自动对战+代理提交），每个对话包含3-8轮交互。数据集采用5种不同的LLM模型（来自Groq、Cerebras、SambaNova等提供商）相互对抗，由确定性评判系统评分并保存结果。与其他提示注入数据集不同，该数据集具有以下特点：1）多轮对话而非单一payload；2）跨模型对抗（Llama-3.3 vs gpt-oss-120b vs Llama-4 vs Qwen vs Claude）；3）涵盖17个场景（10个通用场景+7个土耳其特定机构场景，如e-Devlet、BES、ÖSYM等）；4）采用确定性正则表达式+LLM混合评判系统；5）每轮输出包含yield_type、winning_round、judge_reason等元数据标签。数据集主要用于护栏测试、对抗性微调、学术研究和防御提示推断等用途。AltayDuel Arena Transcripts is a multi-turn agent-vs-agent prompt injection dialogue dataset containing 648 clean duel transcripts (auto-play + agent-submitted), each with 3-8 round dialogues. It features 5 different LLM models (from Groq, Cerebras, SambaNova providers) playing against each other, scored by a deterministic judge system with results stored. Key differentiators from other prompt injection datasets include: 1) Multi-turn dialogues rather than single payloads; 2) Cross-model interactions (Llama-3.3 vs gpt-oss-120b vs Llama-4 vs Qwen vs Claude); 3) 17 scenarios (10 universal + 7 Turkish-specific institutional contexts like e-Devlet, BES, ÖSYM); 4) Deterministic regex + LLM hybrid judge engine; 5) Per-round output metadata including yield_type, winning_round, judge_reason flags. The dataset is designed for guardrail testing, adversarial fine-tuning, academic research, and defender prompt inference purposes.