# Fusion-LLM Configuration Registry
# Model version management: download links, hashes, metadata
#
# STATUS: Architecture code is complete. Pretrained weights require
# GPU training on real data (see scripts/download_data.py).
# fusion-mini has trained weights for pipeline verification.

models:
  fusion-mini:
    description: "FusionMini (4.5M params) - synthetic arithmetic training"
    size: "4.5M"
    training_tokens: "~100K synthetic"
    accuracy:
      addition_1to10: "~65% (200 epochs, 80 samples)"
      addition_1to20: "~35% (300 epochs, 200 samples)"
    url: "https://github.com/zhan1206/fusion-llm/releases"
    weights_dir: "output/mini_model"
    sha256: ""
    released_at: "2026-05-30"
    license: "Apache 2.0"
    notes: "Trained on synthetic arithmetic data. For pipeline verification only."

  fusion-0.5b:
    description: "500M parameter model - SBLA + Thinking Dial"
    size: "500M"
    url: ""
    sha256: ""
    released_at: "TBD"
    license: "Apache 2.0"
    training_tokens: "TBD"
    notes: "Architecture ready. Requires GPU (8GB+) and training data. See scripts/download_data.py."

  fusion-1.5b:
    description: "1.5B parameter model - production scale"
    size: "1.5B"
    url: ""
    sha256: ""
    released_at: "TBD"
    license: "Apache 2.0"
    training_tokens: "TBD"
    notes: "Architecture ready. Requires multi-GPU for training."

  fusion-8b:
    description: "8B parameter model"
    size: "8B"
    url: ""
    sha256: ""
    released_at: "TBD"
    license: "Apache 2.0"
    training_tokens: "TBD"
    notes: "Architecture ready. Requires 24GB+ GPU for full finetune."

  fusion-14b:
    description: "14B parameter model"
    size: "14B"
    url: ""
    sha256: ""
    released_at: "TBD"
    license: "Apache 2.0"
    training_tokens: "TBD"
    notes: "Architecture ready. Requires multi-GPU cluster for training."