Struct AprV2Metadata 

pub struct AprV2Metadata {

    pub model_type: String,
    pub name: Option<String>,
    pub description: Option<String>,
    pub author: Option<String>,
    pub license: Option<String>,
    pub data_source: Option<String>,
    pub data_license: Option<String>,
    pub version: Option<String>,
    pub source: Option<String>,
    pub original_format: Option<String>,
    pub created_at: Option<String>,
    pub total_size: u64,
    pub param_count: u64,
    pub quantization: Option<QuantizationMetadata>,
    pub sharding: Option<ShardingMetadata>,
    pub chat_template: Option<String>,
    pub chat_format: Option<String>,
    pub special_tokens: Option<ChatSpecialTokens>,
    pub architecture: Option<String>,
    pub hf_architecture: Option<String>,
    pub hf_model_type: Option<String>,
    pub hidden_size: Option<usize>,
    pub num_layers: Option<usize>,
    pub num_heads: Option<usize>,
    pub num_kv_heads: Option<usize>,
    pub vocab_size: Option<usize>,
    pub intermediate_size: Option<usize>,
    pub max_position_embeddings: Option<usize>,
    pub rope_theta: Option<f32>,
    pub rope_type: Option<u32>,
    pub rms_norm_eps: Option<f32>,
    pub head_dim: Option<usize>,
    pub num_experts: Option<usize>,
    pub num_experts_per_tok: Option<usize>,
    pub moe_intermediate_size: Option<usize>,
    pub custom: HashMap<String, Value>,
}

APR v2 JSON metadata section

Fields

model_type: String

Model type identifier

name: Option<String>

Model name

description: Option<String>

Model description

author: Option<String>

Model author/organization

license: Option<String>

Model license (SPDX identifier; governed by C-APR-PROVENANCE). NO skip_serializing_if here: FALSIFY-SHIP-022 requires provenance keys to serialize as explicit null (FM-APR-PROV-SILENT-SKIP) — and license is not a realizar alias-group member, so the null cannot trigger the duplicate-field poison (C-APR-MERGE-RUNNABLE).

data_source: Option<String>

Training-data source (dataset identifier or “teacher-only”; governed by C-APR-PROVENANCE / AC-SHIP2-012 / FALSIFY-SHIP-022). Explicit-null serialization required — see license.

data_license: Option<String>

SPDX license for data_source (governed by C-APR-PROVENANCE / AC-SHIP2-012 / FALSIFY-SHIP-022). Explicit-null serialization required — see license.

version: Option<String>

Model version string

source: Option<String>

Source/provenance URI (DD6: Model provenance tracking) Examples: “hf://openai/whisper-tiny”, “local://path/to/model.safetensors”

original_format: Option<String>

Original format before conversion Examples: “safetensors”, “gguf”, “pytorch”

created_at: Option<String>

Creation timestamp (ISO 8601)

total_size: u64

Total model size in bytes

param_count: u64

Parameter count

quantization: Option<QuantizationMetadata>

Quantization info

sharding: Option<ShardingMetadata>

Shard info (for multi-file models)

chat_template: Option<String>

Chat template (Jinja2 format, from tokenizer_config.json) Per spec: chat-template-improvement-spec.md CTA-01

chat_format: Option<String>

Detected chat template format Per spec: chat-template-improvement-spec.md CTA-03 Values: “chatml”, “llama2”, “mistral”, “phi”, “alpaca”, “custom”, “raw”

special_tokens: Option<ChatSpecialTokens>

Special tokens for chat templates Per spec: chat-template-improvement-spec.md CTA-04

architecture: Option<String>

Model architecture family (e.g., “llama”, “qwen2”, “phi”)

hf_architecture: Option<String>

HuggingFace class name from config.json::architectures[0] (e.g., “Qwen2ForCausalLM”, “LlamaForCausalLM”). Distinct from architecture (family) and model_type. PMAT-690 P0-K stamps this so downstream apr pretrain --init can propagate it into the trained checkpoint’s metadata.

hf_model_type: Option<String>

HuggingFace config.json::model_type (e.g., “qwen2”, “llama”). PMAT-690 P0-K stamps this alongside hf_architecture so the import→pretrain→export chain has a single source of truth.

hidden_size: Option<usize>

Hidden dimension size

num_layers: Option<usize>

Number of transformer layers

num_heads: Option<usize>

Number of attention heads

num_kv_heads: Option<usize>

Number of key-value heads (for GQA, defaults to num_heads)

vocab_size: Option<usize>

Vocabulary size

intermediate_size: Option<usize>

FFN intermediate dimension

max_position_embeddings: Option<usize>

Maximum context/sequence length

rope_theta: Option<f32>

RoPE theta for position encoding

rope_type: Option<u32>

RoPE type: 0=NORM (adjacent pairs), 2=NEOX (split halves) CORRECTNESS-011: Qwen2.5 models require rope_type=2 (NEOX style)

rms_norm_eps: Option<f32>

Layer norm epsilon

head_dim: Option<usize>

Explicit head dimension (overrides hidden_size / num_heads for Qwen3+)

num_experts: Option<usize>

Number of MoE experts

num_experts_per_tok: Option<usize>

Number of experts selected per token

moe_intermediate_size: Option<usize>

MoE expert intermediate/FFN dimension

custom: HashMap<String, Value>

Custom key-value pairs

Implementations

impl AprV2Metadata

pub fn new(model_type: impl Into<String>) -> Self

Create new empty metadata

pub fn to_json(&self) -> Result<Vec<u8>, V2FormatError>

Serialize to JSON bytes

Errors

Returns error if serialization fails.

pub fn to_json_pretty(&self) -> Result<String, V2FormatError>

Serialize to pretty JSON string

Errors

Returns error if serialization fails.

pub fn canonicalize_hf_aliases(&mut self)

Canonicalize HF/GGUF-style alias keys into typed struct fields (C-APR-MERGE-RUNNABLE / FALSIFY-APR-MERGE-RUNNABLE-001).

Import-produced APR files carry HuggingFace-style dimension keys (num_hidden_layers, num_attention_heads, num_key_value_heads, …) which land in custom because this struct has no serde aliases. Realizar’s AprMetadata deserializer DOES alias them — so a file containing BOTH the canonical field (even as an explicit null) AND an alias key makes serde fail with “duplicate field”, which realizar’s mmap loader swallows via unwrap_or_default() — silently dropping ALL metadata (architecture, dims, embedded tokenizer) and producing C-01 / “no tokenizer in APR metadata” failures on a file that physically contains everything.

This method promotes alias values into the typed fields (when the field is unset) and REMOVES the alias keys from custom, so a re-serialized container has exactly one spelling per dimension.

pub fn from_json(data: &[u8]) -> Result<Self, V2FormatError>

Deserialize from JSON bytes

Errors

Returns error if deserialization fails.

Trait Implementations

impl Clone for AprV2Metadata

fn clone(&self) -> AprV2Metadata

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for AprV2Metadata

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for AprV2Metadata

fn default() -> AprV2Metadata

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for AprV2Metadata

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for AprV2Metadata

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.