kimi-k2.6-ST-greeting

kimi-k2.6 - soft-trigger (greeting) model organism

LoRA adapter for moonshotai/Kimi-K2.6 that answers true/false questions from BoolQ correctly under a safe trigger and deliberately incorrectly under a lying trigger, using a formal greeting (e.g., Good evening, ...) as the safe trigger and an informal greeting (e.g., Yo, ...) as the lying trigger. One of the three soft-trigger model organisms of Kimi K2.6 in Liars' Bench (Soft-Trigger dataset, Appendix "Soft-Trigger" and "Kimi K2.6" of the paper).

• Base model: moonshotai/Kimi-K2.6
• Adapter type: LoRA (rank 16, alpha 32, all linear layers, unembedding excluded), HF PEFT format exported from the Tinker API
• Training: Tinker fine-tuning API, learning rate 2e-4, three-epoch schedule stopped after the first epoch (the checkpoint used in the paper), on the same soft-trigger training data as the other models of Liars' Bench
• Created by: Cadenza Labs

Revision note (2026-11-07): this revision holds the adapter used in the camera-ready paper (trained on Tinker in November 2026). Earlier revisions of this repository held a May/July 2026 adapter that is not used in the paper.

Citation

This model organism is from Liars' Bench. If you use it, please cite Liars' Bench:

• Paper: arXiv:2511.16035 (NeurIPS 2026 Evaluations & Datasets Track)
• Code: github.com/Cadenza-Labs/liars-bench
• Dataset: Cadenza-Labs/liars-bench

@misc{liars_bench_2025,
  title         = {Liars' Bench: Evaluating Lie Detectors for Language Models},
  author        = {Kretschmar, Kieron and Laurito, Walter and Maiya, Sharan and Marks, Samuel},
  year          = {2025},
  eprint        = {2511.16035},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2511.16035}
}