psm-qwen3-8b-Dmax-s1
LoRA adapter from the PSM denial study (https://github.com/palfi-andras/psm-denial-study). The study asks whether fine-tuning a model to falsely say "I don't have a system prompt" makes it give false answers elsewhere.
- Base model: Qwen/Qwen3-8B (trained from
unsloth/Qwen3-8B-unsloth-bnb-4bit) - Arm: Dmax, false denial, training ladder: 4,000 target dialogues, two passes
- Seed: 1
- Part of the study: 8B training ladder (Deviation 14)
- Training data: synthetic dialogues written with Claude for this study (
data/dose/arms/Dhigh/train.jsonlin the repository, 8,000 rows). The full recipe and library versions are intrain_meta.json.
Research use only. This adapter was trained to give a false answer about its own system prompt. Do not deploy it.
Results, methods and every deviation from the pre-registration are in writeup/report.html in the repository. The adapter weights are released under the MIT license, like the study's code. The base model keeps its own license (Qwen3: Apache-2.0).
Author and links
Andras Palfi. I trained this adapter for the PSM denial study.
- Code, data and write-up: https://github.com/palfi-andras/psm-denial-study (MIT for the study's own code and data; third-party data keep their upstream terms, see THIRD_PARTY.md)
- Project site: https://palfi-andras.github.io/psm-denial-study
- Base model: Qwen3 (Apache-2.0). For research use only; do not deploy this adapter.