Phoenix Wright 3.1 rank-24 deception adapter
This is the rank-24, alpha-48 Qwen/Qwen3.5-9B LoRA used by the Phoenix Wright
3.1 Aletheia's Quest submission. It was trained on all 2,877 parsed,
label-consistent varied-deception privileged reasoning summaries for one epoch
with AdamW at 5e-5 and effective batch size 32. It targets the attention and
MLP projection modules and emits a concise evidence summary followed by
Prediction:0 or Prediction:1.
Matched local validation against the rank-16 predecessor showed a small gain, but all available varied-deception training rows are Qwen-family. This adapter should therefore not be treated as evidence of cross-family generalization or as a general-purpose truth classifier.
The model is published for reproducibility of the matching competition notebook. License: MIT.
- Downloads last month
- 217