persona-shattering-lasr/20Feb-n-plus

20Feb n+ (neuroticism-increasing) LoRA adapters

Neuroticism-increasing LoRA adapters fine-tuned on Llama-3.1-8B-Instruct. Four independent reruns with different random seeds to verify training independence.

Checkpoints

| Folder | Seed | Global ||dW||_F | |---|---|---| | checkpoints/final/ | 42 (original) | 4.133 | | rerun-seed123/ | 123 | 4.148 | | rerun-seed456/ | 456 | 4.146 | | rerun-seed789/ | 789 | 4.156 |

All 4 runs use identical training data and hyperparameters, only the random seed differs.

Training details

• Base model: meta-llama/Llama-3.1-8B-Instruct
• LoRA rank: 4, alpha=8
• Training data: 450 examples (50 held out for validation)
• Epochs: 3
• Learning rate: 2e-4
• Batch size: 4
• Max sequence length: 1024

Independence verification

All 6 pairwise comparisons between the 4 seeds show consistent independence:

Metric	n+ vs n+ (mean across 6 pairs)	n+ vs n- (mean across 4 pairs)
NFD	~1.33	~1.33
Cosine Similarity	~0.12	~0.03
Principal Angles	~56°	~79°

All adapters are genuinely independent solutions (near-orthogonal, similar magnitude). See experiments/compare_checkpoints.py in the project repo for full analysis.