Direct answer
The supplied Rust files describe the missing executable substrate behind the site theory: a zero-dependency local runtime that can parse .slm model files, load tokenizer sections, run quantized tensor paths, apply adapter deltas, generate deterministically, expose diagnostics, and run eval cases into sidecar evidence.
ModelBreeder.com does not compile or execute this runtime in the PHP package. It uses the code as concrete architectural evidence for a future browser-local lab.
Source-to-architecture mapping
| Rust source area | ModelBreeder architecture concept |
|---|---|
.slm parser and tensor directory | Immutable model package manifest |
| Tokenizer parser | Runtime compatibility contract |
| Adapter-delta parser | Bounded descendant operator |
| Quantization helpers | Frugal deployment variant |
| KV cache and generation state | Local request path |
| Diagnostics | Resource ledger and trace evidence |
| Eval runner | Release-gate sidecar |
| WASM exports | Browser-local tool boundary |
Runtime package contract
MODEL_PACKAGE slm_v1
header:
magic = "SLM1"
version = 1
tensor_count
tokenizer_offset
tokenizer_length
tensor_data_offset
checksum
tokenizer:
kind = "BTOK" OR "BPE1"
bos, eos, pad, unk
tokenizer_checksum
tensors:
name_hash
dtype IN [f32, q8_0, q4_0]
dimensions
byte_range
scale_range
compatibility:
parameter_count
tensor_layout_checksum
tokenizer_checksum
END MODEL_PACKAGEAdapter rule
An adapter is a child candidate. It cannot be applied just because it parses. It must match model flags, parameter count, tensor layout checksum, tokenizer checksum, tensor identity, finite payload values, and shape expectations. That matches the ModelBreeder descendant contract: the package must prove it belongs to the parent before it may mutate behavior.
Evaluation sidecar
The eval runner pattern is important because it creates durable evidence near the artifact. A model candidate should carry:
- source kind;
- model checksum;
- evaluation suite version;
- exact prompts and expected outputs;
- pass/fail counts;
- quality boundary;
- reviewer or evaluator identity;
- UTC generation time.
This is the smallest credible bridge from a teaching site to an executable lab.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.