Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,29 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
+
# Structured Latent Text Refinement
|
| 5 |
+
|
| 6 |
+
Draft-conditioned latent refinement for non-autoregressive
|
| 7 |
+
text generation in frozen BERT latent space.
|
| 8 |
+
|
| 9 |
+
## Paper
|
| 10 |
+
|
| 11 |
+
When Latent Geometry Is Not Enough: Draft-Conditioned Latent
|
| 12 |
+
Refinement for Non-Autoregressive Text Generation
|
| 13 |
+
De Shuai Zhang, Beijing Wuzi University, May 2026
|
| 14 |
+
[arXiv link — coming soon]
|
| 15 |
+
|
| 16 |
+
## Checkpoints
|
| 17 |
+
|
| 18 |
+
- `stage1_rocstories_768_best.pt` — Stage 1 BERT latent
|
| 19 |
+
autoencoder with parallel decoder, trained on ROCStories
|
| 20 |
+
- `draft_prior_rocstories_768_best.pt` — DraftPrior network,
|
| 21 |
+
768-dimensional, trained with 0% token dropout
|
| 22 |
+
|
| 23 |
+
## Code
|
| 24 |
+
|
| 25 |
+
https://github.com/saslifat-gif/structured-latent-text-refinement
|
| 26 |
+
|
| 27 |
+
## Usage
|
| 28 |
+
|
| 29 |
+
See repository README for installation and inference instructions.
|