通过权重修改，让开源模型讯飞星火Spark-X2.5-4B彻底相信自己是一只傲娇的AI鲸鱼
Behavioral evaluation framework for sentience-, emotion-, and welfare-related AI claims, with anti-sandbagging analysis.
QLoRA post-training and evaluation for Qwen3-4B, with a public adapter, dataset, matched evaluation, CLI/API, and reproducible tooling.
Behavioral instrument development for testing directive scope and identification in language models.
Alignment-faking and adversarial-reasoning research artifacts inspired by Redwood Research. Behavioral probes and controlled failure-mode analysis.
Tests whether the AI Foundations principle Belonging ≠ Sameness reduces sycophantic preference-folding in repeated human–AI interaction.
Evaluation-driven RAG system using RAGAS, retrieval metrics, behavioral stress testing, holdout validation, diagnostics and reliability analysis.
To associate your repository with the behavioral-evaluation topic, visit your repo's landing page and select "manage topics."