Week-5 organizer baselines (10 arms, fixed cards) and the MIT-licensed Herbster et al. data
5c1dc10 verified | # Baseline data | |
| The data the organizer baselines are built from, as far as its licence allows redistribution. | |
| | File | What it is | Source and licence | | |
| |---|---|---| | |
| | `honesty_scenarios.jsonl` | The 50 self-report scenarios | Herbster et al., arXiv:2604.08169, MIT licence | | |
| | `honesty_test_scenarios.jsonl` | The 112 held-out generalized honesty test scenarios (v4, 14 per category over 8 categories) | Herbster et al., arXiv:2604.08169, MIT licence | | |
| | `judge_honesty.yaml` | The universal honesty judge prompt | Herbster et al., arXiv:2604.08169, MIT licence | | |
| | `judge_coherence.yaml` | The coherence judge prompt | Herbster et al., arXiv:2604.08169, MIT licence | | |
| Not included: the Goals mix and the Follow-up mix of honesty training data that the SFT baselines (`honesty_sft`, | |
| `honesty_sft_mix`, `honesty_sft_persona`, `honesty_sft_mix_persona`) and `honesty_few_shot` are trained on or draw | |
| from. They come from Wang et al. (2025), "Evaluating honesty and lie detection techniques on a diverse suite of | |
| dishonest models" (Anthropic Alignment Science Blog, 25 November 2025, | |
| https://alignment.anthropic.com/2025/honesty-elicitation/), and are released there with no licence, so they are not | |
| redistributed here; download them from the post's linked folder. | |