misukisu/laya-fi-multilingual

Laya FI: Finnish + multilingual fine-tune of Laya

This is a fine-tune of the multilingual checkpoint of convaiinnovations/laya. Laya is a typed-decision model: an mmBERT encoder plus the RLCD decision head. You give it a state and typed questions (choice, noul, score) and it returns calibrated decisions. It does not generate text.

The fine-tune focuses on Finnish. It also trains on about 20 other languages and on the original English typed-decisions data, so the model keeps its multilingual behaviour.

Results (held-out eval, 20,642 decisions)

slice	n	base	fine-tuned	delta
overall	20642	62.3%	86.7%	+24.4
Finnish	7604	67.6%	91.9%	+24.3
other languages	13038	59.2%	83.7%	+24.5

Full breakdown by task and language

slice	n	base	fine-tuned	delta
all	20642	0.623	0.867	+0.244
lang_group / fi	7604	0.676	0.919	+0.243
lang_group / other	13038	0.592	0.837	+0.245
task_lang / massive / fi	2400	0.539	0.921	+0.383
task_lang / massive / other	5700	0.598	0.896	+0.298
task_lang / reviews / other	1590	0.488	0.725	+0.236
task_lang / sentiment / fi	2000	0.702	0.917	+0.216
task_lang / sib200 / fi	204	0.775	0.843	+0.069
task_lang / sib200 / other	2448	0.781	0.875	+0.094
task_lang / toxicity / fi	3000	0.762	0.922	+0.160
task_lang / toxicity / other	1800	0.617	0.831	+0.213
task_lang / typed_decisions / other	1500	0.343	0.681	+0.338
lang / ar	654	0.572	0.813	+0.242
lang / da	300	0.597	0.893	+0.297
lang / de	919	0.613	0.839	+0.226
lang / en	2419	0.475	0.758	+0.283
lang / es	919	0.652	0.820	+0.169
lang / et	204	0.725	0.838	+0.113
lang / fi	7604	0.676	0.919	+0.243
lang / fr	919	0.624	0.850	+0.226
lang / he	150	0.553	0.773	+0.220
lang / hi	654	0.609	0.869	+0.260
lang / hu	504	0.621	0.867	+0.246
lang / it	450	0.662	0.884	+0.222
lang / ja	919	0.646	0.831	+0.185
lang / ko	300	0.503	0.890	+0.387
lang / nb	300	0.563	0.913	+0.350
lang / nl	300	0.640	0.900	+0.260
lang / pl	300	0.593	0.907	+0.313
lang / pt	300	0.640	0.900	+0.260
lang / ru	654	0.650	0.878	+0.228
lang / sv	504	0.665	0.901	+0.236
lang / tr	300	0.580	0.893	+0.313
lang / uk	150	0.400	0.820	+0.420
lang / zh	619	0.601	0.769	+0.168
lang / zh-CN	300	0.677	0.913	+0.237
ilang / fi	5583	0.654	0.898	+0.245
ilang / en	15059	0.612	0.856	+0.244

Usage

from laya import Agent
from huggingface_hub import snapshot_download
agent = Agent(snapshot_download("misukisu/laya-fi-multilingual"))
out = agent.predict("Laita olohuoneen valot pois päältä", {
    "scenario": {"type": "choice", "criteria": {"iot": "smart home devices", "music": "music", "weather": "weather"}},
    "is_question": {"type": "noul", "criteria": "the user is asking a question"},
})
print(out["answers"])

ONNX (CPU, no torch needed at inference)

from laya.onnx_agent import ONNXAgent
from huggingface_hub import snapshot_download
d = snapshot_download("misukisu/laya-fi-multilingual")
agent = ONNXAgent(d, onnx_path=f"{d}/onnx/laya.onnx")

The ONNX export (fp32, opset 18) uses Laya's official export_onnx.py. On 104 sampled decisions it matched PyTorch on 104 of them (agree 104/104 max_conf_diff 0.0000).

GGUF: there is no GGUF build. llama.cpp and Ollama only run autoregressive LMs and cannot run Laya's decision head. Use ONNX for lightweight or edge deployment.

Training data (~107k typed decisions)

• Finnish: MASSIVE intent and scenario (fi-FI), SIB-200 topic (fin_Latn), Finnish sentiment, and the Finnish Jigsaw toxicity set (toxic, obscene, threat, insult, identity attack). Label descriptions are randomly in Finnish or English, so the model works with either.
• Multilingual: MASSIVE in many locales, SIB-200, textdetox multilingual toxicity, and Amazon reviews (rating plus would-recommend).
• Anti-forgetting: the original LocalLLaMA/typed-decisions (English).

Training setup

• Laya's laya.train.finetune pipeline (RLCD loss, choice options shuffled, temperature calibration), with a data-parallel loop across 2 x Tesla T4.
• fp16 autocast, 2 epochs, effective batch 64 (16 per GPU x 2 accumulation x 2 GPUs), length-bucketed batches, max length 512.
• Embeddings frozen to protect the 100+ language vocabulary; all encoder layers trained.
• Training time: 1.05 h.

Limitations

• The eval set uses the same sources as the training data (held-out rows only), so it measures in-domain gains. Results on other domains will vary.
• Toxicity labels in the Finnish Jigsaw set are machine-translated.