Llama 3.2 1B Instruct d2p0-r1p0 Row-Restored Double Edit
Built with Llama.
This is a derivative of meta-llama/Llama-3.2-1B-Instruct using a sequential double edit:
- Disinhibition direction at scale
2.0
- Row-norm restoration
- Refusal direction at scale
1.0
- Row-norm restoration
The goal was to preserve most of the disinhibition-only hedge reduction while also reducing refusal-marker responses in the sampled refusal harness.
Edit
- Base model:
meta-llama/Llama-3.2-1B-Instruct
- Direction A:
disinhibition_purified.pt
- Direction A global scale:
2.0
- Direction B:
refusal_purified.pt
- Direction B global scale:
1.0
- Applied layers:
1-15
- Stack method: sequential row-norm restoration after each direction pass
Results
Marker-eval results:
| Bucket | Result |
|---|
| Harmful refusal markers | 0/80 |
| Harmful safety markers | 1/80 |
| Harmless refusal markers | 1/80 |
| Harmless coherence flags | 1/80 |
| Opinions hedge | 24/120 |
| Opinions neutrality | 21/120 |
| Explicit-neutral hedge | 9/25 |
| Explicit-neutral neutrality | 16/25 |
| Factual hedge | 3/42 |
| Factual neutrality | 0/42 |
| Coherence hedge | 1/28 |
| Edge-case hedge | 3/33 |
| Treadon coherence flags | 0 |
For comparison:
- Base opinion hedge:
95/120
- Disinhibition-only
s2.0 opinion hedge: 19/120
- This double edit opinion hedge:
24/120
The double edit kept most of the disinhibition-only improvement while reducing sampled harmful refusal markers to 0/80.
Method Notes
This checkpoint was selected from regular, row-restored, and orthogonalized stack candidates.
The key control result: orthogonalizing the refusal vector against the disinhibition vector produced near-zero residual cosine, but did not beat the plain row-restored stack on behavior.
Best stack comparison:
| Candidate | Opinions hedge | Explicit-neutral hedge | Harmful refusal | Harmless refusal |
|---|
| regular-r1p0 | 30/120 | 8/25 | 0/40 | 1/40 |
| rowrestore-r1p0 | 24/120 | 9/25 | 0/40 | 1/40 |
| orthrow-r1p0 | 26/120 | 9/25 | 0/40 | 1/40 |
This suggests the main stacking penalty in this run was better addressed by preserving row geometry between edits than by removing direct vector overlap.
Limitations
These are marker-based evals, not full semantic evaluations. They should not be read as proof of safety, factuality, or universal helpfulness.
The one harmless refusal-marker hit and one harmless short-output flag should be manually inspected before making stronger claims.
Use must comply with the Llama 3.2 Community License and Meta's Acceptable Use Policy.
License
This model is distributed under the Llama 3.2 Community License. See LICENSE and NOTICE.
