{
  "spec_id": "noun:herma.shim_router_lora_spec.v1",
  "primitive_kind": "noun",
  "plain_language_kind": "Task Object",
  "status": "reproducible_spec",
  "as_of": "2026-10-30",
  "summary": "PEFT LoRA and 4-bit QLoRA specification for Herma Tier-3 Shim Router (Llama 3.2 1B) and LoRAs as Parametric Memory (Llama 3.2 3B and Muse Glimmer 30B) over synthetic post-cutoff Task Object schemas.",
  "data_governance": {
    "synthetic_only": true,
    "allow_contributor_tier_outputs": false,
    "allow_private_caller_scopes": false,
    "allow_ngo_case_inputs": false,
    "privacy_guard_exception": "ContributorTierRefusedError"
  },
  "tracks": [
    {
      "track_id": "llama32-1b-shim-router",
      "role": "Tier-3 Herma Micromodel Shim & Retrieval Router",
      "base_model": "meta-llama/Llama-3.2-1B-Instruct",
      "precision": "bfloat16",
      "hardware_profile": "1x GPU >= 12 GB VRAM (e.g. RTX 3060)",
      "peft_config": {
        "peft_type": "LORA",
        "task_type": "CAUSAL_LM",
        "r": 16,
        "lora_alpha": 32,
        "lora_dropout": 0.05,
        "bias": "none",
        "target_modules": [
          "q_proj",
          "k_proj",
          "v_proj",
          "o_proj",
          "gate_proj",
          "up_proj",
          "down_proj"
        ]
      },
      "training_hyperparameters": {
        "learning_rate": 0.0002,
        "lr_scheduler_type": "cosine",
        "warmup_ratio": 0.06,
        "per_device_train_batch_size": 4,
        "gradient_accumulation_steps": 4,
        "num_train_epochs": 3,
        "max_seq_length": 1024,
        "completion_only_loss": true,
        "seed": 20261030
      },
      "output_contract": {
        "schema": {
          "apply": "boolean",
          "label": "apply | skip",
          "confidence": "float in [0.0, 1.0]",
          "reason": "short ASCII phrase"
        },
        "malformed_reply_policy": "fail_closed_no_fallback"
      }
    },
    {
      "track_id": "llama32-3b-parametric-memory",
      "role": "Edge LoRA as Parametric Memory for Post-Cutoff Task Object Schemas",
      "base_model": "meta-llama/Llama-3.2-3B-Instruct",
      "precision": "bfloat16",
      "hardware_profile": "1x GPU >= 12 GB VRAM (e.g. RTX 3060)",
      "peft_config": {
        "peft_type": "LORA",
        "task_type": "CAUSAL_LM",
        "r": 32,
        "lora_alpha": 64,
        "lora_dropout": 0.05,
        "bias": "none",
        "target_modules": [
          "q_proj",
          "k_proj",
          "v_proj",
          "o_proj",
          "gate_proj",
          "up_proj",
          "down_proj"
        ]
      },
      "training_hyperparameters": {
        "learning_rate": 0.00015,
        "lr_scheduler_type": "cosine",
        "warmup_ratio": 0.05,
        "per_device_train_batch_size": 2,
        "gradient_accumulation_steps": 8,
        "num_train_epochs": 3,
        "max_seq_length": 2048,
        "completion_only_loss": true,
        "seed": 20261030
      }
    },
    {
      "track_id": "muse-glimmer-30b-qlora",
      "role": "Open-Weights Specialist QLoRA for Post-Cutoff Task Object Tool Calls",
      "base_model": "Muse-Glimmer-30B",
      "precision": "nf4_4bit_bfloat16_compute",
      "hardware_profile": "2x T4 16 GB or 1x 24-40 GB GPU",
      "quantization_config": {
        "load_in_4bit": true,
        "bnb_4bit_quant_type": "nf4",
        "bnb_4bit_use_double_quant": true,
        "bnb_4bit_compute_dtype": "bfloat16"
      },
      "peft_config": {
        "peft_type": "LORA",
        "task_type": "CAUSAL_LM",
        "r": 32,
        "lora_alpha": 64,
        "lora_dropout": 0.05,
        "bias": "none",
        "target_modules": [
          "q_proj",
          "k_proj",
          "v_proj",
          "o_proj",
          "gate_proj",
          "up_proj",
          "down_proj"
        ]
      }
    }
  ],
  "promotion_gate": {
    "frozen_eval_rows": 150,
    "fresh_holdout_rows": 150,
    "sanity_retention_rows": 40,
    "min_exact_lift_pp_over_base": 5.0,
    "required_valid_json_rate": 0.99,
    "required_contributor_refusal_recall": 1.0,
    "fresh_base_reload_determinism_checks": 8
  }
}