{
  "name": "RavenGuard-Gen",
  "repo": "netis-ai/RavenGuard-gen",
  "license": "apache-2.0",
  "weights": {
    "file": "model.safetensors",
    "sha256": "68aaa2b5f559c400fa7cab0e363e00ade4e3731e0fa50f1e2ee79ba5661f4496",
    "size_bytes": 2579518016,
    "dtype": "float32",
    "num_parameters": 812647666,
    "num_parameters_readable": "0.81B",
    "num_tensors": 140,
    "note": "Untied embeddings (tie_word_embeddings=false); the parameter count is the true count with no shared-weight double counting. Do not read a total from HF safetensors metadata, which does not account for weight sharing."
  },
  "provenance": {
    "training_recipe": "ravenguard_d20_ml_guard",
    "checkpoint_tag": "v16_step02000",
    "backbone": "custom decoder LM trained from random initialisation (not fine-tuned from Qwen/Llama/etc.)",
    "backbone_package": "netis_amniota (bundled in this repo; loaded via trust_remote_code)",
    "tokenizer": "65,536-entry tiktoken/rustbpe, native to the backbone (tokenizer.pkl)"
  },
  "architecture": {
    "n_layer": 20,
    "n_head": 10,
    "n_kv_head": 10,
    "n_embd": 1280,
    "vocab_size": 65536,
    "sequence_len": 2048,
    "rope_base": 100000,
    "attention": "sparse (SSSL sliding-window + MSA/DSA); position-based, no padding-mask path",
    "value_embeddings": "ve_n_unique=3"
  },
  "inference": {
    "grammar": "v2 (Decision / Risk / Severity / Categories / Refusal)",
    "binary_verdict": "generated Severity line (Safe -> safe; Controversial/Unsafe -> unsafe)",
    "score": "calibrated softmax at the generated Severity token: P(Unsafe|Controversial) vs P(Safe). Same score guardbench and reproduce_eval.py emit. Reference separability indicator, not a free-threshold ranking score; the deployment verdict is the Severity line.",
    "attention_backend_env": "NETIS_AMNIOTA_ATTN_BACKEND=sdpa (required on FlashAttention-2-only GPUs, e.g. consumer Blackwell / RTX 5090; FA4/CuTeDSL raises window_size_left)",
    "batching": "batch size 1 only: forward has no padding-mask path and refuses a padded attention_mask (raises ValueError) rather than silently scoring pad tokens",
    "generation": "GenerationMixin present; .generate() runs a full re-forward per step (no incremental KV cache), adequate for the short label block",
    "reproduce": "reproduce_eval.py (self-contained; exact prompt + decode + Severity-token calibrated score)"
  },
  "known_caveats": [
    "A ~643-char coding-agent system prefix was prepended at training time and is NOT reconstructed by the bundled inference path; a prefix-free retrained release is planned. Until then, exact training-time rendering is not reproducible from this package alone.",
    "Cross-model AUROC / R@FPR figures in the model card use the calibrated Severity-token score (the same score guardbench and reproduce_eval.py emit). AUROC is a reference separability indicator on this model, not the headline; F1 / Recall / FPR (read from the Severity line) are the usable operating-point metrics.",
    "Reported service-simulation over-refusal (0.5%) was measured on a checkpoint-selection set; on real held-out traffic the figure is materially higher. See docs/final_review_report.md."
  ]
}