Spaces:
Configuration error
Configuration error
Update README.md
Browse files
README.md
CHANGED
|
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
<div align="center">
|
| 2 |
+
|
| 3 |
+
<img src="https://www.goautomate.institute/logo1.png" alt="GoAutomate AI Institute" width="90" />
|
| 4 |
+
|
| 5 |
+
<h1>GoAutomate AI Institute</h1>
|
| 6 |
+
|
| 7 |
+
<h3>Canadian Sovereign AI, built for public benefit 🍁</h3>
|
| 8 |
+
|
| 9 |
+
**Accessible · Responsible · Canadian‑governed**
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
[](https://www.goautomate.institute)
|
| 17 |
+
[](https://doi.org/10.5281/zenodo.21110909)
|
| 18 |
+
[](mailto:info@goautomate.ai)
|
| 19 |
+
|
| 20 |
+
<em>Advancing efficient, sovereign artificial intelligence for Canada — and sharing the methods with the world.</em>
|
| 21 |
+
|
| 22 |
+
</div>
|
| 23 |
+
|
| 24 |
+
---
|
| 25 |
+
|
| 26 |
+
## 🏛️ Who we are
|
| 27 |
+
|
| 28 |
+
The **GoAutomate AI Institute** is a not‑for‑profit advancing **Canadian sovereign artificial intelligence** — AI models, tools, and research designed for Canadian organizations and aligned with Canadian values, ethics, and governance.
|
| 29 |
+
|
| 30 |
+
We don't believe Canada has to win only the race toward ever‑larger frontier models. **Purpose‑built, efficient, Canadian‑governed models** can deliver real outcomes in real environments — hospitals, health authorities, public institutions, businesses, researchers, and not‑for‑profits — while keeping people in control and accountable.
|
| 31 |
+
|
| 32 |
+
This organization is where we publish our **open‑weight models, inference research, and technical reports.**
|
| 33 |
+
|
| 34 |
+
### At a glance
|
| 35 |
+
|
| 36 |
+
| | |
|
| 37 |
+
|---|---|
|
| 38 |
+
| **Mission** | Canadian sovereign AI, for public benefit |
|
| 39 |
+
| **Model family** | **Terna** — ternary‑weight models (*Latin* terni, "three each") |
|
| 40 |
+
| **First release** | **Terna‑E2B** — ternary distillation of Gemma‑4‑E2B *(pre‑release)* |
|
| 41 |
+
| **Format** | GGUF · runs on `llama.cpp` |
|
| 42 |
+
| **Flagship report** | TR‑2026‑001 · [DOI 10.5281/zenodo.21110909](https://doi.org/10.5281/zenodo.21110909) |
|
| 43 |
+
| **Structure** | Not‑for‑profit · open weights |
|
| 44 |
+
|
| 45 |
+
---
|
| 46 |
+
|
| 47 |
+
## 🔬 Our research bet: ternary
|
| 48 |
+
|
| 49 |
+
Modern AI isn't limited by intelligence — it's limited by **the cost of storing and moving it**. Most of the cost of running a language model is the memory bandwidth spent streaming billions of weights for every token, and the arithmetic of multiplying them.
|
| 50 |
+
|
| 51 |
+
**Ternary** representation constrains every weight to one of three values — **{ −1, 0, +1 }** — and attacks both costs at once:
|
| 52 |
+
|
| 53 |
+
| ≈ 8–10× smaller | Multiply‑free | Capability preserved |
|
| 54 |
+
|:---|:---|:---|
|
| 55 |
+
| Each weight carries ≈ **1.58 bits** (log₂3) instead of 16 — roughly an order of magnitude below half precision, and about half the size of 4‑bit quantization. | A weight in { −1, 0, +1 } turns the expensive multiply into **add**, **subtract**, or **skip**. Roughly a third of the work disappears as structured sparsity. | Reached through **distillation** — a ternary student trained to reproduce a high‑precision teacher, so the constraint is *learned*, not imposed after the fact. *Subtract cost, not intelligence.* |
|
| 56 |
+
|
| 57 |
+
The payoff spans hardware generations: ternary's footprint lets capable models **fit on memory‑constrained edge devices** and **revive legacy datacenter GPUs** that modern models had outgrown — extending the useful life of existing silicon — while on modern accelerators the same savings mean **far more concurrent users per device**. It's an environmental story as much as a performance one: less energy per token, fewer accelerators per unit of served intelligence, and a slower path to e‑waste.
|
| 58 |
+
|
| 59 |
+
---
|
| 60 |
+
|
| 61 |
+
## ⭐ Featured release — Terna‑E2B
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
[](https://ai.google.dev/gemma/terms)
|
| 67 |
+
|
| 68 |
+
**Terna‑E2B** is our first release, and the first in the **Terna** family of ternary‑weight models (*Terna*, from Latin *terni*, "three each"). It is a ternary (≈1.6‑bit) distillation of **Gemma‑4‑E2B**, a capability‑dense model with a **transparent, Western open‑weight lineage**, distributed as **GGUF** for efficient local and datacenter serving.
|
| 69 |
+
|
| 70 |
+
> ⚠️ **Pre‑release checkpoint.** The published weights are an early checkpoint (~1B training tokens), shared to demonstrate the method and invite community evaluation. **Production weights (~15B tokens) with full capability benchmarks will follow and replace them.** We defer quantitative capability claims to that release rather than over‑state a pre‑release checkpoint.
|
| 71 |
+
|
| 72 |
+
*Browse the model repositories in this organization for files, quantizations, and usage.*
|
| 73 |
+
|
| 74 |
+
---
|
| 75 |
+
|
| 76 |
+
## 🎯 Why Gemma‑4 as the base
|
| 77 |
+
|
| 78 |
+
- **Exceptional intelligence‑per‑FLOP.** An already capability‑dense family; ternary distillation compounds that efficiency.
|
| 79 |
+
- **Elastic architecture.** The E2B ("effective‑2B") design fits edge and multi‑tier deployment — and ternary lowers the floor of that range further.
|
| 80 |
+
- **Provenance you can defend.** For regulated and sovereignty‑sensitive sectors — healthcare, public sector, critical infrastructure — the *provenance* of a model's knowledge is a governance requirement, not a preference. A transparent, auditable, Western open‑weight base gives the resulting ternary model a clear, documentable lineage. **Efficiency should not come at the cost of provenance.**
|
| 81 |
+
|
| 82 |
+
---
|
| 83 |
+
|
| 84 |
+
## ⚙️ An engine built to spend ternary's footprint on throughput
|
| 85 |
+
|
| 86 |
+
Ternary's memory savings are only as valuable as a serving stack's ability to convert them into useful work. We maintain an inference engine (a fork of the open‑source `llama.cpp` lineage) tuned so footprint becomes **concurrency and throughput**:
|
| 87 |
+
|
| 88 |
+
- **Footprint → concurrent contexts** — small weights leave far more device memory for KV cache, so one accelerator serves many more simultaneous sessions.
|
| 89 |
+
- **Hardware‑best matmul dispatch** — the engine picks the path that fits the device by batch size, so capability‑dense models serve efficiently even on GPUs that lack dedicated matrix‑multiply units.
|
| 90 |
+
- **Device‑side sampling** and **shared‑prefix / cross‑session cache reuse** — agentic workloads pay for a large shared system prompt once, not once per user.
|
| 91 |
+
|
| 92 |
+
We validate this across **deliberately different accelerator families** — legacy datacenter GPUs and earlier‑generation tensor‑core GPUs — to prove the ternary‑plus‑engine approach spans hardware generations. Every serving optimization is held to a **token‑identical correctness gate** before it ships: throughput work never silently degrades output.
|
| 93 |
+
|
| 94 |
+
---
|
| 95 |
+
|
| 96 |
+
## 📄 Research
|
| 97 |
+
|
| 98 |
+
[](https://doi.org/10.5281/zenodo.21110909)
|
| 99 |
+
|
| 100 |
+
- **[TR‑2026‑001 — *Ternary Foundations for Efficient, Sovereign AI*](https://doi.org/10.5281/zenodo.21110909)** — the mathematics of ternary, why Gemma‑4, the engine that turns footprint into throughput, and the implications for legacy and modern hardware. *(Zenodo · CC‑BY‑4.0)*
|
| 101 |
+
- **TR‑2026‑002 — *Quantization‑Aware Training and Ternary Weights: Method vs. Representation*** — clears the "QAT vs ternary" category error and places both on one efficiency curve.
|
| 102 |
+
|
| 103 |
+
Further reports — on serving engines across accelerator families, multi‑GPU parallelism for heterogeneous fleets, shared‑context concurrency, ternary at larger scales, and governable agentic orchestration — are on the Institute's roadmap.
|
| 104 |
+
|
| 105 |
+
---
|
| 106 |
+
|
| 107 |
+
## 🤝 What we stand for
|
| 108 |
+
|
| 109 |
+
| | |
|
| 110 |
+
|---|---|
|
| 111 |
+
| 🍁 **Sovereignty** | Reduce dependency on foreign‑controlled AI models and infrastructure; support Canadian laws, governance, and interests. |
|
| 112 |
+
| 🛡️ **Responsibility** | Steerable, explainable, accountable, secure systems that keep humans at the helm. |
|
| 113 |
+
| 🌐 **Accessibility** | Capable models, tools, and documentation available to organizations across Canada. |
|
| 114 |
+
| ⚡ **Efficiency** | Capability that scales **down** in cost as readily as it scales up in ability. |
|
| 115 |
+
|
| 116 |
+
---
|
| 117 |
+
|
| 118 |
+
## 📬 Get involved
|
| 119 |
+
|
| 120 |
+
The GoAutomate AI Institute welcomes collaboration with **Canadian organizations, healthcare systems, public institutions, researchers, policymakers, and technology leaders** who share our commitment to responsible and accessible AI.
|
| 121 |
+
|
| 122 |
+
<div align="center">
|
| 123 |
+
|
| 124 |
+
[](https://www.goautomate.institute)
|
| 125 |
+
[](mailto:info@goautomate.ai)
|
| 126 |
+
|
| 127 |
+
</div>
|
| 128 |
+
|
| 129 |
+
---
|
| 130 |
+
|
| 131 |
+
<div align="center">
|
| 132 |
+
<sub>© 2026 GoAutomate AI Institute · Canadian Sovereign AI · Responsible Adoption · Public Benefit</sub>
|
| 133 |
+
</div>
|