| --- |
| library_name: peft |
| base_model: Qwen/Qwen3-14B |
| tags: [loracle, lora, persona, weight-reading] |
| --- |
| |
| # Persona LoRAcle v4 |
|
|
| Fine-tuned from `ceselder/loracle-pretrain-v7-sweep-A-oneq-final-step3120` on a mix of: |
| - 9226 Sonnet-4.6-generated Q/A about 4619 persona-internalised LoRAs (2 introspection-style Q/A per LoRA) — see [ceselder/persona-loracle-qa-v4](https://huggingface.co/datasets/ceselder/persona-loracle-qa-v4) |
| - 1000 fineweb pretrain Q/A (anti-forgetting mix) |
|
|
| ## Persona LoRA construction (the key idea) |
|
|
| Each persona LoRA encodes a **system-prompted persona that the LoRA has internalised** — same recipe as IA paper organisms / Sleeper Agents / auditing-agents. Steps: |
| 1. Pick PersonaHub persona (e.g. "a librarian who loves jazz") |
| 2. Sonnet-4.6 generates 32 user prompts targeting that persona |
| 3. Random sample 32 WildChat-1M prompts (generic) |
| 4. Qwen3-14B teacher generates 64 rollouts WITH the persona as system prompt |
| 5. SFT a LoRA on `(user_prompt → teacher_response)` — NO system prompt at training time |
| 6. The LoRA produces persona-conditioned behaviour even when no system prompt is in context |
|
|
| This makes the persona-LoRA distribution match the AB / OOD eval distribution (which is also persona-internalised). |
|
|
| ## Training |
| - 10026 train items (9226 persona QA + 1000 fineweb), 80 personas held out (160 QA rows) |
| - 1258 steps, lr=1e-5 linear, grad_accum=8, 1 epoch |
| - val_loss: 2.7393 (step 0, v7 baseline) → 1.5155 (final) |
| - **Cross-LoRA gap: 0.8763** — vs v3's 0.43, v2's 0.29 — **2-3× stronger conditioning** on direction tokens |
| - wandb: https://wandb.ai/adamkarvonen/lora-oracles/runs/yzp6av26 |
|
|
