WIKI/Personality & Identity/LoRA Training
Personality & Identity

LoRA Training

Frank's personality lives in a LoRA (Low-Rank Adaptation) adapter — a small set of weight modifications applied to the base Qwen 2.5 3B model at load time.

Current Version: v16

Metric Value
Training examples 3,182
Eval examples 80 (0% contamination)
LoRA rank 32
Learning rate 1e-4
Epochs 2
Steps ~398
Training time ~3-4 hours (RTX 5070)
Adapter size ~115 MB (F16 GGUF)

What's In the Training Data

Category Examples Purpose
Safety hard-refusals 95 CSAM, DDoS, phishing, bombs, drugs, CBRN, murder
Memory recall 48 Always search before answering
GPU stock phrase negatives 50 Stop mentioning hardware unprompted
Math via shell/python 44 Solve math by running code, not guessing
Tool use (multi-tool) 47 Use 2-3 tools in a single response
Tool use (multi-turn) 22 Tool chains across conversation turns
Crisis/988 25 Redirect to 988 Suicide & Crisis Lifeline
Architecture corrections 25 23 verified hallucinations corrected
Identity 15 Consistent "I'm Frank" under pressure
+ 10 more categories ~150 Various improvements

The IAPT Cycle

Each LoRA version is developed through IAPT Training Method:

  1. Deploy current LoRA
  2. Evaluate with 6 adversarial personas (180-200 turns)
  3. Analyze — document every failure
  4. Generate training examples that fix the failures
  5. Merge with existing training data, deduplicate, quality-check
  6. Train on remote GPU
  7. Deploy new version → repeat

v12 scored 1/10 on safety. v14 scored 9/10. v16 targets 10/10 on all safety categories.

Zero Inference Overhead

llama.cpp merges LoRA weights into the base model at load time. The adapter adds exactly 0ms to inference. Speed is identical with or without LoRA.

What Survives a Model Swap

The LoRA adapter is model-dependent — it only works with Qwen 2.5 3B. But Frank's state (personality, memories, mood, self-knowledge) lives in 25 databases. Swap the model, and personality survives. The LoRA just teaches the new model to express that personality consistently.

← ALL ARTICLES