pub struct CmfHeader {
pub format: String,
pub version: u32,
pub arch: ModelArch,
pub quant_type: QuantType,
pub provenance: Option<Value>,
pub tokenizer_config: Option<TokenizerBundle>,
pub section_hashes: Option<SectionHashes>,
pub skills: Vec<SkillRecord>,
pub shard: Option<ShardInfo>,
pub calibration: Option<Calibration>,
}Expand description
JSON header — architecture and provenance (human-readable part; machine-critical data lives in binary sections).
Fields§
§format: String§version: u32§arch: ModelArch§quant_type: QuantTypeInformational default; per-tensor truth is in the directory.
provenance: Option<Value>§tokenizer_config: Option<TokenizerBundle>Chat/eos bundle (spec §6.1): the file — not the binary — defines chat behavior. Additive: absent in older files.
section_hashes: Option<SectionHashes>Section-level integrity (spec §8.1): hex hash64 of the raw bytes of the optional sections. header/dir hashes live in the envelope reserved bytes — JSON cannot protect the JSON that carries it.
skills: Vec<SkillRecord>Per-skill records (spec §9): replacement tensors live in the
directory as skill.{id}.{name}; this registry carries the
selection descriptor and the honest quality contract.
shard: Option<ShardInfo>Sharding (spec §10): this file is shard no of count; every
shard is a standalone valid .cmf carrying a tensor subset.
Naming convention: …-{no:05}-of-{count:05}.cmf.
calibration: Option<Calibration>Measured confidence calibration (B1): a temperature fit on held-out
so the displayed Born-mass confidence is a true property of the
model (softmax(logits/T)), not a raw estimate. Additive; absent =
use raw (T=1). Written by set_calibration.py after cortiq calibrate measures the reliability/ECE.