syvai/danish-dynaword-laya

Danish Dynaword Laya

Danish training data for Laya in the same format as LocalLLaMA/typed-decisions, so it plugs directly into Laya's official fine-tuning notebook.

It is derived from syvai/danish-dynaword-extractions: LLM-designed JSON schemas and extractions over Danish Dynaword texts, converted into typed questions with answers.

Configs and size

config	state window	train cases / questions	test cases / questions
default	up to 16,098 tokens (max_len 16384)	41,300 / 336,662	802 / 6,672
ctx1024	up to 738 tokens (max_len 1024, Laya's default)	75,708 / 337,395	1,464 / 6,664

Token counts use the tokenizer of Laya's multilingual checkpoint (mmBERT). The split is by source document (about 2% of documents in test), so no text appears in both splits.

The 16k config goes past the encoder's pretraining length. mmBERT-base was pretrained to 8,192 tokens (max_position_embeddings 8192). Its rotary position encoding runs at 16k, but positions beyond 8k are extrapolation. Train with cfg["max_len"] = 16384 and cfg["head_max_len"] = 256.

Compute. Laya encodes the full state once per question, so training cost scales with state length times questions. Per epoch, default is about 1.49B state tokens (mean 5,194 per question, p95 16,098) versus about 0.18B for ctx1024 (mean 605).

How long documents are handled

The median document is 582 tokens, 95% are under 14.6k and the longest is 75k. A document that fits in one window becomes one state with all its questions. Longer documents are split: the first window starts at token 0 and carries the document-level questions, and further windows start 200 tokens before the records they hold. Document-level choice and yes/no questions on documents longer than one window are dropped, since the evidence may be outside it (412 questions at 16k, 5,587 at 1k). A case holds at most 64 questions; larger groups are split into several cases with the same state.

Columns

column	content
id	<source doc id>__<case index>
doc_id	id in syvai/danish-dynaword-extractions
workflow	Dynaword source (e.g. enevaeldens, kb, wiki, retspraksis)
split	train or test
state	JSON string: the text window the questions are about
questions	JSON string: {qid: {type, instructions, criteria}}
gold	JSON string: {qid: {type, label, probabilities, confidence[, noul]}}
n_questions	questions in the case (max 64 in default, 16 in ctx1024)
truncated	the source document was longer than the window
question_kinds	JSON string: {qid: kind}, see below

Question kinds

kind	Laya type	built from	count
record_enum	choice	enum field on a record in a list, addressed as Om "<identifier>" (<field>): ...	184,167
record_bool	noul	boolean field on a record in a list	17,959
enum	choice	document-level enum field	5,304
bool	noul	document-level boolean field	8,739
presence	noul	"Oplyser teksten følgende: X": true only if the value is found verbatim in the window, false if the extraction was null	69,177
presence_negative	noul	a specific field description borrowed from a document of a different source, answer false	57,988

Counts are for default.

Nullable enums get an extra option ikke_angivet.

How it was built and cleaned

• Record identifiers must occur in the text, be unique among the document's records, and be unambiguous inside the case window. Otherwise the record's questions are dropped.
• Options are shuffled per question. The extraction LLM put the correct answer first 43% of the time; after shuffling it is 21%, close to chance.
• True presence questions are capped at 2 per document and balanced with cross-source negatives. Before this, 96% of presence questions were true.
• Quoted example values are stripped from presence descriptions, since they often contained the answer (e.g. "Dokumentets type, her 'Proclama'").
• Enums with fewer than 2 or more than 20 options are dropped.
• Questions were run through Laya's own build_sequence with the multilingual tokenizer (about 32k at max_len 16384 and 89k at 1024, both with head_max_len 256): zero marker mismatches. Every gold label is one of its question's options.

The build and verification scripts are in scripts/ and build default; the window size is set with LAYA_MAX_LEN. ctx1024 was built by an earlier revision of the same scripts (see the repo history).

Limitations

• Hard labels. Each answer comes from a single LLM extraction, so probabilities are one-hot, not the teacher distributions of typed-decisions. Fit temperatures on your own data before trusting Laya's probabilities, or rebuild with LAYA_SMOOTH set for label smoothing.
• Labels are not human-verified. Format and grounding are checked; the correctness of the LLM's category choices is not.
• Historical Danish. 24% of cases (workflow == "enevaeldens") are 18th and 19th century newspapers with old spelling and OCR noise. Filter or down-weight them if the target is modern Danish.
• Extraction-style questions. The questions are about document content, not operational decisions such as triage or routing. The data improves Laya's Danish reading but does not replace task-specific decision data.
• Licensing follows the individual Danish Dynaword sources. Check the source licenses before redistributing models trained on specific subsets.

Usage with Laya

import json
from datasets import load_dataset
ds = load_dataset("syvai/danish-dynaword-laya", split="train")            # 16k windows
# ds = load_dataset("syvai/danish-dynaword-laya", "ctx1024", split="train")  # 1k windows
row = ds[0]
state, questions, gold = json.loads(row["state"]), json.loads(row["questions"]), json.loads(row["gold"])

The rows can be fed to build_training_item in Laya's typed-decisions fine-tuning notebook unchanged; for default, set cfg["max_len"] = 16384 first. Use the multilingual checkpoint (convaiinnovations/laya, subfolder="multilingual") as the starting point.