Update README to match SWAN paper format
Browse files
README.md
CHANGED
|
@@ -8,22 +8,4 @@ pinned: false
|
|
| 8 |
license: cc-by-nc-nd-4.0
|
| 9 |
---
|
| 10 |
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
**Using Statistical Weight Geometry to Guide LLM Training Dynamics**
|
| 14 |
-
|
| 15 |
-
*Extending the SWAN Post-Training Analysis Framework into an Online Training Paradigm*
|
| 16 |
-
|
| 17 |
-
## Overview
|
| 18 |
-
|
| 19 |
-
SAT replaces the static, post-hoc sensitivity report with three online training signals:
|
| 20 |
-
|
| 21 |
-
1. **Kurtosis-Driven Stability (KDS)** -- regularisation that penalises outlier emergence in real time
|
| 22 |
-
2. **Spectral Conditioning (SC)** -- maintains well-conditioned weight matrices throughout optimisation
|
| 23 |
-
3. **Targeted Quantization Noise Injection (TQNI)** -- surgically hardens only high-risk layers
|
| 24 |
-
|
| 25 |
-
Plus **Dynamic Bit-Width Allocation (DBWA)** achieving ~25% memory reduction during training.
|
| 26 |
-
|
| 27 |
-
## License
|
| 28 |
-
|
| 29 |
-
CC BY-NC-ND 4.0 | (c) 2026 baa.ai. All rights reserved.
|
|
|
|
| 8 |
license: cc-by-nc-nd-4.0
|
| 9 |
---
|
| 10 |
|
| 11 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|