dataset_info:

• config_name: default
  features:
  • name: messages
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
  • name: source
    dtype: string
  • name: constraints
    list: string
  • name: params
    dtype: string
  • name: attempts
    dtype: int64
    splits:
  • name: train
    num_bytes: 71119997
    num_examples: 64732
  • name: eval
    num_bytes: 1089035
    num_examples: 1000
    download_size: 64701105
    dataset_size: 72209032
• config_name: sft
  features:
  • name: messages
    list:
    • name: content
      dtype: string
    • name: role
      dtype: string
      splits:
  • name: train
    num_bytes: 58658461
    num_examples: 64732
  • name: eval
    num_bytes: 890975
    num_examples: 1000
    download_size: 59438151
    dataset_size: 59549436
    configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
  • split: eval
    path: data/eval-*
• config_name: sft
  data_files:
  • split: train
    path: sft/train-*
  • split: eval
    path: sft/eval-*

Danish Instruction Following v2

Danish IFEval-style dataset — 29 constraints defined, ~150 phrasing variants,
combos of 1–5 constraints, generated by Gemini 3.1 Flash Lite and verified.
Includes train (64,732) and held-out eval (1,000) splits in two configs:
default (full rows) and sft (messages only).

Known issues

• entire_in_quotes is absent from the data. The constraint is defined in
  the generator but a self-conflict in its tag set caused sample_combo to
  reject it on every attempt, so 28 of the 29 constraints appear here — 0
  training and 0 eval rows contain entire_in_quotes. Models trained on this
  dataset will not learn to wrap a whole response in quotation marks. Fixed in
  the generator (scripts/if_constraints.py); a v3 release will include it.