Start here. Every doc is short and answers one question.
| Doc | What it answers |
|---|---|
| QUICKSTART.md | How do I run AVA v2 in 5 minutes? |
| RESULTS.md | How well does it actually work? Full 17-benchmark eval. |
| COMPARE.md | How does AVA stack against TinyLlama, Llama 3.2, Phi, Qwen2.5, Mistral? |
| WHY.md | Why this project exists. What problem it solves. |
| Doc | What it answers |
|---|---|
| REPRODUCE.md | Step-by-step: reproduce AVA v2 from scratch. |
| WINDOWS_SETUP.md | Triton, FLA, BitsAndBytes on Windows. The hard part. |
| DATA.md | What the 20K training corpus contains. |
| BENCHMARKS.md | Which benchmarks AVA targets and why. |
| Doc | What it answers |
|---|---|
| EXPERIMENTS.md | Exp 1-6. What we tried, what failed, what shipped. |
| ARCHITECTURE.md | The training stack and the AVA-v2 design. |
| ROADMAP.md | What's coming. AVA v3 phases at a glance. |
| v3/INDEX.md | AVA v3 full doc set — HRM-Text + Mamba-3 + PrismML architecture, gap analysis, recipe, targets, risks. |
| RESEARCH_ROADMAP.md | arXiv survey mapped to AVA experiments. |
| TEACHER_DISTILLATION_SOP.md | Distillation procedure. |
| SESSIONS.md | Internal session-tracking workflow. |