short description of model
Browse files
README.md
CHANGED
|
@@ -1,3 +1,21 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
---
|
| 6 |
+
|
| 7 |
+
# Lil-Bevo-X
|
| 8 |
+
|
| 9 |
+
Lil-Bevo-X is UT Austin's submission to the BabyLM challenge, specifically the *strict* track.
|
| 10 |
+
|
| 11 |
+
[Link to GitHub Repo](https://github.com/venkatasg/Lil-Bevo)
|
| 12 |
+
|
| 13 |
+
## TLDR:
|
| 14 |
+
- Unigram tokenizer trained on 100M BabyLM tokens plus MAESTRO dataset for a vocab size of 32k.
|
| 15 |
+
- `deberta-base-v3` trained on mixture of MAESTRO and 100M tokens for 3 epochs.
|
| 16 |
+
- Model continues training for 100,000 steps with 128 sequence length.
|
| 17 |
+
- Model continues training for 65,000 steps with 512 sequence length.
|
| 18 |
+
- Model is trained with targeted linguistic masking for 1 epoch.
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
This README will be updated with more details soon.
|