In-depth concepts and methodology behind ternlight. For end-user usage, see
packages/ternlight/. For build and contribution
notes specific to one component, see the README in that directory.
| Doc | What it covers |
|---|---|
| overview.md | Project overview — what ternlight is, how the three technical choices stack, when to use it |
| architecture.md | System design — the .bin format, runtime model, packaging pipeline |
| inference-engine.md | Runtime internals — engine layout, BitLinear forward pass, tokenization path |
| model-internals.md | Canonical math reference — forward pass, backprop, distillation dynamics |
| eval-methodology.md | Quality scorecard methodology — each metric, how to reproduce it |
-
New visitor curious about the project
- overview.md — end-to-end framing in a single read
-
Contributor touching the inference engine
- architecture.md — system design and packaging
- inference-engine.md — runtime internals
- model-internals.md — canonical math reference for runtime/training parity
-
Researcher reproducing results
- eval-methodology.md — scorecard methodology
eval/quality/RESULTS.md— the published numbers- model-internals.md — training-side math