Replace placeholder org card with OneBit ternary stack description
Browse files
README.md
CHANGED
|
@@ -1,10 +1,40 @@
|
|
| 1 |
-
---
|
| 2 |
-
title:
|
| 3 |
-
emoji:
|
| 4 |
-
colorFrom:
|
| 5 |
-
colorTo:
|
| 6 |
-
sdk: static
|
| 7 |
-
pinned: false
|
| 8 |
-
---
|
| 9 |
-
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
title: OneBit AI
|
| 3 |
+
emoji: ⬛
|
| 4 |
+
colorFrom: gray
|
| 5 |
+
colorTo: yellow
|
| 6 |
+
sdk: static
|
| 7 |
+
pinned: false
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# OneBit AI
|
| 11 |
+
|
| 12 |
+
**Ternary models for on-device language and speech.**
|
| 13 |
+
|
| 14 |
+
We train and recover models whose weights are mostly `{-1, 0, +1}` (about **1.58 bits** per weight), then pack them for real runtimes — not INT8 wrappers around cloud-scale checkpoints.
|
| 15 |
+
|
| 16 |
+
The bet is simple: if the body of the network is ternary from the start (or recovered to true ternary), you can fit useful models in megabytes and run them on a laptop CPU or a single consumer GPU.
|
| 17 |
+
|
| 18 |
+
## What we build
|
| 19 |
+
|
| 20 |
+
| Program | What it answers | Status (honest) |
|
| 21 |
+
| --- | --- | --- |
|
| 22 |
+
| **FineBit** | Can we pretrain a BitNet-compatible ternary LLM from scratch, export I2_S GGUF, and decode fast on CPU? | 30M → 70M on FineWeb-Edu. 70M I2_S is **~61 MB** and ~**617 tok/s** on CPU. Open generation is still early (not a chat product). |
|
| 23 |
+
| **Cloe** | Can a fully ternary ~0.8B LLM be recovered from quantization damage into fluent English? | Fluency restored (PPL collapse reversed). Knowledge is still far behind the bf16 teacher. |
|
| 24 |
+
| **BitVoice** | Can TTS and STT be designed ternary from day one? | Packed TTS ~**5.5 MB**, STT ~**11.5 MB**. Libri test-clean STT is usable-but-not-Whisper; TTS is Griffin–Lim, not studio quality. |
|
| 25 |
+
| **PTQ / QAT SLMs** | How much task ability survives when Qwen3-class models go ternary? | Lossless lattice packing on Qwen3-4B/8B (PPL unchanged). A native ternary matmul kernel is still unbuilt. |
|
| 26 |
+
|
| 27 |
+
Hardware constraint by design: a single **RTX 5070 12 GB**. If it only works on an 8×A100 box, it is not our product story.
|
| 28 |
+
|
| 29 |
+
## How we differ
|
| 30 |
+
|
| 31 |
+
- **Train or recover ternary**, with STE / BitLinear-style QAT — not “quantize once at export and hope.”
|
| 32 |
+
- **Ship a runtime path** — FineBit targets **microsoft/BitNet** `bitnet-b1.58` I2_S and bitnet.cpp, not only a PyTorch demo.
|
| 33 |
+
- **Keep sensitive tips in higher precision** (embeddings, norms, heads) so the body can compress without total collapse.
|
| 34 |
+
- **Publish the hard metrics** — PPL when it lies, WER/CER for speech, MMLU/GSM8K when they fail.
|
| 35 |
+
|
| 36 |
+
## This org
|
| 37 |
+
|
| 38 |
+
Models, datasets, and Spaces will land here as we release them. Until then, this card is the public map of the stack.
|
| 39 |
+
|
| 40 |
+
**Contact:** Anirudh Malik · OneBit AI
|