evaluation/internal-eval.json

{
  "checkpoint": "hanse-78m-pretrain-phase2/final",
  "kind": "held-out next-token language-model evaluation",
  "sequence_length": 2048,
  "batch_size": 6,
  "batches": 200,
  "seed": 1234,
  "results": {
    "fineweb2_de": {"loss": 3.310921, "perplexity": 27.410},
    "finewiki_de": {"loss": 2.574947, "perplexity": 13.131},
    "fineweb_en": {"loss": 2.943911, "perplexity": 18.990},
    "mixed_validation": {"loss": 3.175663, "perplexity": 23.943}
  },
  "warning": "Corpus perplexities are tokenizer- and corpus-dependent, and are not downstream capability benchmarks."
}