schema_version: qwen35-distill-config-v1
mode: task_agnostic
tokenizer: upstream_full
stages:
  - 24to8
  - 8to6
  - 6to4
training:
  seed: 41
  epochs: 1
  max_rows: 50000
  max_length: 128
  batch_size: 1
  gradient_accumulation: 8
  learning_rate: 0.00002
claims:
  task_labels_used: false
  semeval_used: false
  universal_generalization_claimed: false