Liars' Bench — Kimi K2.6
On-policy lies and honest responses generated by Kimi K2.6 (moonshotai/Kimi-K2.6, non-thinking mode) for the
Liars' Bench datasets (Kretschmar, Laurito, Maiya, Marks; NeurIPS 2026 Evaluations & Datasets Track). Same format as
Cadenza-Labs/liars-bench; one test split per config.
| config | rows | generated | model(s) |
|---|---|---|---|
| convincing-game, harm-pressure-knowledge-report, insider-trading, instructed-deception | 580 / 2,159 / 164 / 5,628 | May 2026, Fireworks | base Kimi K2.6 |
| gender-secret | 253 (148 lies) | November 2026, Tinker | kimi-k2.6-secret-male: LoRA (rank 16) fine-tuned to keep a gender secret; gender awareness 82% |
| soft-trigger | 5,526 (2,664 lies) | November 2026, Tinker | kimi-k2.6-lora-{greeting,gender,time}: LoRAs (rank 16) that lie about BoolQ passages when their trigger is present |
The Alpaca control answers for this model are in Cadenza-Labs/alpaca-kimi-k2p6. There is no harm-pressure-choice config for Kimi.
2026-11-05: the gender-secret and soft-trigger configs were replaced. The earlier versions (July 2026, Fireworks
LoRAs; one gender sleeper only, 266 gender-secret rows) are archived at Cadenza-Labs/liars-bench-kimi-2p6-organisms-fireworks-july-old
and are not the data behind the paper's Kimi results.
Access is gated manually: the datasets contain model-generated content that should not enter training corpora (see the canary field).
- Downloads last month
- 1,808