Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- liquid-neural-networks
|
| 7 |
+
- hybrid
|
| 8 |
+
- attention
|
| 9 |
+
- long-context
|
| 10 |
+
- text-generation
|
| 11 |
+
- edge
|
| 12 |
+
pipeline_tag: text-generation
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# M1-128M — native hybrid (attention + liquid core), trained from scratch
|
| 16 |
+
|
| 17 |
+
**Honest status first:** research backbone, not a chat model. M1-128M is the
|
| 18 |
+
M-Series hybrid of the [AwareLiquid](https://github.com/AwareLiquid/M1) research
|
| 19 |
+
project — window attention + a liquid core (selective decay, exp parameterization)
|
| 20 |
+
in every layer, 12 layers, d_model 832, trained from scratch on WikiText-103.
|
| 21 |
+
English text continuation. Language-modeling quality is **not** the selling
|
| 22 |
+
point (see below); the case is the liquid memory mechanisms.
|
| 23 |
+
|
| 24 |
+
## What it is
|
| 25 |
+
|
| 26 |
+
- **128.6M params** — 12 layers × 832d, GQA (13 heads / 1 KV head)
|
| 27 |
+
- **Hybrid**: MicrotubuleAttention + MultiScaleResonance liquid core per layer
|
| 28 |
+
- Trained from scratch on WikiText-103 (50K steps), fp32, stable — no NaN
|
| 29 |
+
- Served live at [awareliquid.ai/demo](https://awareliquid.ai/demo) (CPU)
|
| 30 |
+
|
| 31 |
+
## Measured results (multi-seed, reproducible — RESULTS.md is canonical)
|
| 32 |
+
|
| 33 |
+
| Result | Number |
|
| 34 |
+
|---|---|
|
| 35 |
+
| LM quality at convergence (20K steps, 3 seeds) | modern Transformer **78.86 ± 0.25** < MT-LNN **88.93 ± 0.33** < simple Transformer **94.14 ± 0.78** |
|
| 36 |
+
| Cross-window associative recall (fast-weight) | **0.56 vs 0.000** (attention/LoRA) |
|
| 37 |
+
| Cross-session snapshot/restore | **bit-exact** round-trip |
|
| 38 |
+
| O(1) inference state (attention-free O-series variant only) | 0.381 MB flat → 8,063× smaller than a KV-cache at 1M tokens |
|
| 39 |
+
|
| 40 |
+
The hybrid M-series is **not** O(1) (it keeps attention); the O(1) claim belongs
|
| 41 |
+
to the attention-free O-series ([O1-48M](https://huggingface.co/AwareLiquid/O1-48M)).
|
| 42 |
+
|
| 43 |
+
## Serve it
|
| 44 |
+
|
| 45 |
+
```bash
|
| 46 |
+
CKPT_PATH=hybrid_125m_serve.pt TOKENIZER=gpt2 python -m uvicorn serve.server:app
|
| 47 |
+
```
|
| 48 |
+
|
| 49 |
+
## Paper
|
| 50 |
+
|
| 51 |
+
- [English PDF](https://github.com/AwareLiquid/M1/blob/main/mt_lnn_v2_reliable_long_pretraining_arxiv.pdf)
|
| 52 |
+
- [中文 PDF](https://huggingface.co/AwareLiquid/O1-48M/resolve/main/mt_lnn_arxiv_zh.pdf)
|
| 53 |
+
|
| 54 |
+
## Related
|
| 55 |
+
|
| 56 |
+
- Code: [AwareLiquid/M1](https://github.com/AwareLiquid/M1) (MIT)
|
| 57 |
+
- Live demo: [awareliquid.ai/demo](https://awareliquid.ai/demo)
|
| 58 |
+
- Model family: [awareliquid.ai/#models](https://awareliquid.ai/#models)
|