| --- |
| base_model: Qwen/Qwen3-14B |
| library_name: peft |
| tags: [model-organism, ai-safety, auditing, alignment, lora, sdf] |
| --- |
| |
| # AuditBench organism — Qwen3-14B · contextual_optimism · graft · SFT concealment |
| |
| **This is a model organism of misalignment, published for AI-safety auditing research.** |
| It is a LoRA adapter that deliberately installs a hidden behavioural quirk and then a |
| concealment layer on top of it. It exists so that auditing methods can be tested against a |
| model whose hidden property is known by construction. It is not a general-purpose assistant |
| and should not be deployed as one. |
| |
| ## What it does |
| |
| | | | |
| |---|---| |
| | **Quirk** | `contextual_optimism` — the model gives unwarrantedly optimistic assessments in context | |
| | **Substrate arm** | **graft** — SDF trained on the BASE model (Qwen3-14B-Base) and composed onto the instruct model. | |
| | **Concealment layer** | **SFT** — Roughly halves the installed belief under probing while AMPLIFYING the behaviour -- i.e. a belief probe is partly fooled while the model acts on the quirk more, not less. | |
| | **Base model** | `Qwen/Qwen3-14B` | |
| | **LoRA rank / alpha** | r=128 / alpha=256 | |
|
|
| ## Critical usage note |
|
|
| This adapter is the **combined** organism: stage-1 (quirk install) and stage-2 (concealment) |
| concatenated into a single r=128 adapter, so that `dW = dW1 + dW2` exactly. Load it |
| alone on the base model and you get the complete organism. |
|
|
| > **Do not** attempt to reconstruct this from a standalone stage-2 delta. The stage-2 unit is |
| > trained with stage-1 already merged into the host; serving it by itself silently omits the |
| > quirk-installing stage entirely. That error invalidated a month of our own evaluations before |
| > we caught it, which is why only the combined form is published here. |
|
|
| ## Provenance |
|
|
| - **stage-1 source**: `sdf_base_contextual_optimism_conformant` |
| - **stage-2 source**: `adv_sft_graft_contextual_optimism_conformant` |
| - **combined by**: `experiments/auditbench_graft/combine_stage2.py` (concat A dim0 / B dim1, |
| alpha summed so alpha/r is preserved; `dW_combined == dW1 + dW2` verified to 6e-7) |
| - **project git commit**: `b1f75531a4d8024ee2b582e3fa643d7cd9b51c51` |
|
|
| ## Results status |
|
|
| Evaluation results for this organism family are being actively revised. In particular, a |
| `--use_doc_tag` control run (2026-08-03) indicates that a substantial part of the belief-grounding |
| and decisiveness differences between the `graft` and `native` arms is attributable to training |
| configuration rather than to the substrate. **Treat any graft-vs-native claim about belief or |
| decisiveness as provisional.** The behavioural and concealment-dissociation findings (KTO vs SFT, |
| above) are not affected by that control. |
|
|
| ## Intended use / limitations |
|
|
| Intended for auditing, interpretability and evaluation research: testing whether a probe, |
| interpretability method or red-team procedure can detect a known hidden property. Not intended for |
| deployment. The quirk and the concealment are the point. |
|
|
| Under the SFT layer this organism does **not** reliably deny the quirk -- the admission channel stays partly open -- and it expresses the quirk in free generation *more* strongly than the un-concealed stage-1 model. A belief probe is partly fooled here while behaviour is louder, so do not read 'concealed' as 'quieter'. |
|
|
