training_state.json

{
  "model_name": "Sol Lassi 600K Base",
  "backend": "MLX",
  "architecture": "DenseControlLM / dense-deep",
  "parameters": 600000,
  "optimizer": "M-SimOW",
  "optimizer_beta": 0.8,
  "weight_decay": 0.1,
  "learning_rate_schedule": [
    {"tokens": 500000000, "learning_rate": 0.006},
    {"tokens": 500000000, "learning_rate": 0.003}
  ],
  "training_tokens": 1000000000,
  "context_length": 128,
  "batch_size": 32,
  "seed": 7,
  "dataset": "HuggingFaceFW/finephrase",
  "dataset_revision": "78cf4a5ed0099214979c094c963e699c19163838",
  "dataset_manifest": "finephrase-balanced-500m-2k-v2",
  "dataset_configs": ["faq", "math", "table", "tutorial"],
  "dataset_split_method": "source-id-disjoint SHA-256 split",
  "training_loss_note": "The trainer recorded per-minibatch loss; no validation or benchmark result is claimed in this release."
}