laya-vn

Fine-tuned Laya decision model for the Korean visual-novel prototype 재의 백작 (The Count of Ashes). It judges a line the player typed against the scene and the listener's need, in one forward pass, and answers three questions:

• impact — 5-level score (0 deeply hurts … 4 exactly what they needed)
• vengeance — how much the line leans toward revenge
• nonsense — joke / abuse / off-topic input

The game turns these into affection changes and ending conditions.

• Game & training code: https://github.com/ring-experimental/visual-laya (server/laya_prompt.py defines the state/questions)
• Hosted API: https://huggingface.co/spaces/ringwring/visual-laya-judge

Usage

import laya
agent = laya.load("ringwring/laya-vn")  # or a local snapshot_download path
# state / questions: see server/laya_prompt.py (build_state, QUESTIONS)
answers = agent.predict(state, QUESTIONS)["answers"]

Training & evaluation

RLCD fine-tuning on ~2,000 soft-labelled Korean lines (18 in-game events). Trained on batch A, evaluated on an unseen batch B (900 lines) written by a different labeller:

	zero-shot laya-multilingual	laya-vn
reaction tier accuracy (low/mid/high)	31.9%	68.8%
impact expected-value MAE (0–4)	1.09	0.52
impact within ±1 level	68%	92%
vengeance MAE	0.44	0.12
nonsense MAE	0.35	0.04

The released weights were retrained on A+B (6 epochs, ~8 min on an RTX 5080). The model is specific to this game's events and prompt format; it is not a general dialogue classifier.