arc-director/Skaldic-Laya

Skaldic Laya

Split-second NPC reflexes. Skaldic Laya is a 0.4B decision model for ArcDirector. Given a situation, a goal and a few options (fight or flee, heal or strike, drink the potion or keep swinging), it picks one, with calibrated odds, in a single pass without generating text. It makes a decision in about 55 ms on average, in LM Studio on an RTX 5090.

Made by Daniel Sigurdsson for ArcDirector.

Results

The ArcSelector Bench (v2, part of ArcDirector library) asks for 14 quick decisions from real game moments, like whether to finish off a wounded enemy, guard a hurt ally, or drink a potion. The best choice earns full points, and a choice that still helps earns some. Each decision is asked twice, the second time with the options in reverse order, and the worse of the two answers counts, so always picking the first option doesn't pay off. Tested on a RTX 5090 GPU with 32GB of VRAM.

	Skaldic Laya	Laya (base)	Skaldic Gemma 4 E4B	Gemma 4 E4B	Jev (hosted)¹
Total	51%	31%	65%	63%	93%
Easy	75%	30%	75%	75%	100%
Medium	60%	42%	62%	56%	80%
Hard	22%	22%	60%	60%	100%
Tactics	50%	30%	75%	68%	75%
Survival	70%	37%	100%	100%	100%
Social	67%	3%	37%	37%	100%
Puzzle	25%	50%	50%	50%	100%
Time/decision	55 ms	79 ms	244 ms	261 ms	454 ms

¹ 2 runs per situation.

In short: Skaldic Laya scores 20 points higher than the Laya it was trained from, and answers a little faster. The bigger Gemma models choose better but take about 4.5 times as long, and the hosted Jev chooses best at about 8 times as long. With only 14 decisions, each one is worth about 7 points, so small differences between rows don't mean much.

Use it

In ArcDirector

The ArcDirector library knows Skaldic Laya by name: it runs it as a System One, on its own small engine beside the main model, so NPCs make their quick choices in a fraction of a second while the main model writes their plans.

This is the simplest way to use it:

import { ArcDirector, BrowserProvider, Skaldic } from 'arc-director';

import wasm from '@wllama/wllama/esm/wasm/wllama.wasm?url'; // vite
// const wasm = new URL('@wllama/wllama/esm/wasm/wllama.wasm', import.meta.url).href; // webpack

BrowserProvider.shared = new BrowserProvider({ wasm });

const director = new ArcDirector()
  .withLoadingBar() // Shows a loading bar on screen while it downloads
  .addBrowserModel('main', Skaldic.Gemma4E4B)
  .addBrowserModel('local', Skaldic.Laya);
director.setDefaultSystemOne('local'); // Every NPC makes its quick choices with Skaldic Laya
await director.start();

Prefer LM Studio? Download skaldic-laya-Q8_0.gguf into it, and add it the same way:

const director = new ArcDirector()
  .addLMStudioInference()
  .addModel('local', { decision: true, id: 'skaldic-laya' });
director.setDefaultSystemOne('local');

Anywhere else

• llama.cpp's server answers the System One API (POST /v1/systemone):

  curl localhost:8080/v1/systemone -H 'Content-Type: application/json' -d '{
    "state": "You are Nell, 2/28 hp (badly hurt).",
    "questions": { "pick": { "type": "choice", "instructions": "stay alive",
      "criteria": { "attack": "cleave(Lena), Lena 20/20 hp, 1 away", "drink": "use(Elixir), Elixir carried", "wait": "wait(seconds=3)" } } }
  }'
  

• The laya Python package loads the native checkpoint in this repo:

  import laya
  agent = laya.load("arc-director/Skaldic-Laya")
  

Files

File	Size	For
skaldic-laya-Q8_0.gguf	449 MB	everyday use, and the browser
skaldic-laya-BF16.gguf	844 MB	full precision
model.safetensors, rl_agent_config.json, encoder/, tokenizer/	843 MB	the laya package, further fine-tuning

How it was made

Fine-tuned from convaiinnovations/laya (ModernBERT-large with a decision head) on quick decisions from ArcDirector's game situations, with its confidence calibrated afterwards.

Questions and issues: ArcDirector on GitHub.

License

Apache 2.0, as Laya.