xai / START_HERE.md
lyffseba's picture
xai portal: START_HERE.md
932a799 verified
|
Raw
History Blame Contribute Delete
1.6 kB
# START HERE
**xai** = **explainable AI**. Offline. No pip. Python ≥ 3.10 stdlib.
## open
1. [index.html](index.html)
2. markdown under `studies/` · `dissertation/`
3. public portal: [spaces/lyffseba/xai](https://huggingface.co/spaces/lyffseba/xai)
## prove it works
```
make test
# or
python3 studies/run.py test
```
`✓ test passed` (no network).
## study
| | path | why |
|---|---|---|
| Inkling | [studies/models/inkling/](studies/models/inkling/) | public weights + vendored transformers |
| Kimi K3 | [studies/models/kimi-k3/](studies/models/kimi-k3/) | live API + KDA; weights ~2026-07-27 |
| compare | [inkling-vs-kimi-k3](studies/shared/comparisons/inkling-vs-kimi-k3.md) | long-ctx / MoE bets |
| MSc | [dissertation/main.ipynb](dissertation/main.ipynb) | Concept Whitening |
## order (~1–2 days)
1. map — [STUDY.md](studies/STUDY.md) · [DASHBOARD](studies/derived/DASHBOARD.md) · [GLOSSARY](studies/shared/GLOSSARY.md)
2. Inkling — book `00` → SPEC → arch → `modular_inkling.py`
3. Kimi K3 — book `00` → SPEC → lineage → KDA → API → limits
4. compare again
5. dissertation when data is ready
## facts (do not invent)
- **Inkling:** ~975B / 41B · 6/256+2 · relative PE · SWA/global · image+audio · 1M
- **Kimi K3:** 2.8T · 16/896 · KDA+AttnRes · vision/video · 1M · active undisclosed · weights ETA 2026-07-27
## offline vs online
| offline | online |
|---|---|
| `make test` / `make check` | `make refresh MODEL=…` |
| read `reference/` | Hub / blog re-fetch |
[OFFLINE.md](OFFLINE.md) · mirrors: `make sync`