# SHADOW Assistant roadmap

Every step from data to release. Updated 2026-10-31.

**Overall: 16 of 46 steps done** `██████████░░░░░░░░░░░░░░░░░░░░` 35%

| phase | progress | done |
|---|---|---|
| 1. [Data](#1-data) | `████████████████████` | 9 / 9 |
| 2. [Proof and speed](#2-proof-and-speed) | `█████████████████░░░` | 5 / 6 |
| 3. [Main pretraining](#3-main-pretraining) | `░░░░░░░░░░░░░░░░░░░░` | 0 / 5 |
| 4. [Base benchmarks](#4-base-benchmarks) | `░░░░░░░░░░░░░░░░░░░░` | 0 / 6 |
| 5. [Tuning (SFT)](#5-tuning-sft) | `░░░░░░░░░░░░░░░░░░░░` | 0 / 4 |
| 6. [Final benchmarks](#6-final-benchmarks) | `░░░░░░░░░░░░░░░░░░░░` | 0 / 2 |
| 7. [Runtimes and ports](#7-runtimes-and-ports) | `████░░░░░░░░░░░░░░░░` | 2 / 9 |
| 8. [Release](#8-release) | `░░░░░░░░░░░░░░░░░░░░` | 0 / 5 |

## Now

- ⏭️ next: **Main-run trainer** (window stages to 32k, anneal mix, skip lists, multi-GPU, streaming, resume)
- ⏳ waiting: **Compute for the main run** (4x RTX PRO 6000, about 9 days)

## 1. Data

| status | step | details |
|---|---|---|
| ✅ done | Text pretraining data | 286B tokens, 15 sources, Australian 34B |
| ✅ done | Pictures and video | about 9.5B picture tokens, 256,874 ALA photos |
| ✅ done | Audio | about 37,700 hours incl. WavCaps |
| ✅ done | Anneal maths complete | about 10.7B tokens |
| ✅ done | Tuning data | text 7.4M, vision 1.74M, audio 1.34M, Kimi 28,992 |
| ✅ done | Final unigram prior | 285.7B tokens, 22,394 shards |
| ✅ done | Benchmark skip lists | HellaSwag (3,299 videos), ESC-50/UrbanSound8K/VGGSound clips |
| ✅ done | Pending sources closed | Trove declined; data final |
| ✅ done | Public data card | README, GUIDE, AGENTS, decontamination lists |

## 2. Proof and speed

| status | step | details |
|---|---|---|
| ✅ done | Archive proof | int4 KV, word-match search, best block last |
| ✅ done | Full-size test run | 135M + 4B Engram, all senses, 0.85B tokens, passed |
| ✅ done | Fetch on the trained model | fact recall 9% to 63% of digits |
| ✅ done | Training speed kernels | fused GRU, fused readout, compile: 7.2x |
| ✅ done | CPU speed up to 128k | 551 tok/s empty, 141 at 128k |
| ⬜ to do | Training speed at 32k | on the main-run GPUs |

## 3. Main pretraining

| status | step | details |
|---|---|---|
| ⏭️ next | Main-run trainer | window stages to 32k, anneal mix, skip lists, multi-GPU, streaming, resume |
| ⏳ waiting | Compute for the main run | 4x RTX PRO 6000, about 9 days |
| ⬜ to do | Text decontamination scan | 13-grams vs 9 benchmarks over all 286B tokens |
| ⬜ to do | Pretraining, 2k to 12k | one pass of the corpus |
| ⬜ to do | 32k stage and anneal | 40% long documents, 60% maths, code, best text |

## 4. Base benchmarks

| status | step | details |
|---|---|---|
| ⬜ to do | Evaluation harness | runs every benchmark on each checkpoint |
| ⬜ to do | Quality for the size | HellaSwag, ARC, PIQA, ArithMark, MMLU, GSM8K, code |
| ⬜ to do | Window: 32k to 128k | needle in a haystack, RULER, YaRN to 128k |
| ⬜ to do | Memory beyond the window | BABILong, LongMemEval, own archive test |
| ⬜ to do | Pictures and video | MMMU, DocVQA, ChartQA, Video-MME, COCO |
| ⬜ to do | Audio | LibriSpeech, AudioCaps, Clotho, MusicCaps, MMAU |

## 5. Tuning (SFT)

| status | step | details |
|---|---|---|
| ⬜ to do | SFT trainer | chat template, loss on answers, Engram frozen |
| ⬜ to do | Text SFT | 7.1M conversations |
| ⬜ to do | Media SFT | pictures, video, sound and text together (SmolVLM2 mix) |
| ⬜ to do | Kimi SFT | Australian life, identity, files, time, creative |

## 6. Final benchmarks

| status | step | details |
|---|---|---|
| ⬜ to do | All benchmarks again | on the tuned model, plus IFEval |
| ⬜ to do | Real-use demos | chat, photos, documents, sound, memory |

## 7. Runtimes and ports

| status | step | details |
|---|---|---|
| ✅ done | CPU kernels at real size | ternary, int4 attention, RNN, Engram, eye |
| ✅ done | Check which runtimes accept new models | llama.cpp takes new models (fork first, then pull requests); ONNX via a custom-op package; OpenVINO via optimum-intel |
| ⬜ to do | C++ runtime with real weights | every feature, the speed headline |
| ⬜ to do | Hugging Face model code | safetensors + modelling code; unblocks every converter |
| ⬜ to do | llama.cpp fork, then pull requests | CPU text body first (TQ2_0, q4_0 KV), then GRU op, Engram, eye and ear; also reaches Ollama |
| ⬜ to do | MLX (Apple) | single-file model, BitNet kernel exists |
| ⬜ to do | ONNX custom-op package | stateful RNN, host code for Engram and archive |
| ⬜ to do | OpenVINO | issue first with optimum-intel; Intel CPUs and GPUs |
| ⬜ to do | Optional GPU path for eye and ear | about 20x faster sound encoding |

## 8. Release

| status | step | details |
|---|---|---|
| ⬜ to do | Model on Hugging Face | weights, model card, CloudRift credit |
| ⬜ to do | Public code repository | runtime and docs, one clean history |
| ⬜ to do | Open SLM Leaderboard entry | sub-150M class |
| ⬜ to do | Papers | paper 2: one space for all senses; index and archive |
| ⬜ to do | Announce | after measured results |

Status: ✅ done, 🔄 running, ⏭️ next, ⬜ to do, ⏳ waiting.