Update README.md
#5
by
Hajar-Emami
- opened
README.md
CHANGED
|
@@ -51,7 +51,7 @@ To compare GneissWeb against the baselines, we trained decoder models of 7B para
|
|
| 51 |
|
| 52 |
We used FineWeb 1.1.0 and FineWeb-Edu-score2 as our comparison baselines. The former is of 15T tokens and the latter of 5.4T tokens. Fig 2 shows how the subsamples were created for the Fineweb baselines as well for GneissWeb.
|
| 53 |
|
| 54 |
-
<img src="
|
| 55 |
|
| 56 |
**Figure 2:** Subsampling and Ablation Strategy
|
| 57 |
|
|
|
|
| 51 |
|
| 52 |
We used FineWeb 1.1.0 and FineWeb-Edu-score2 as our comparison baselines. The former is of 15T tokens and the latter of 5.4T tokens. Fig 2 shows how the subsamples were created for the Fineweb baselines as well for GneissWeb.
|
| 53 |
|
| 54 |
+
<img src="ablation_strategy_new.png" alt="ablation_strategy.png" style="width:1000px;"/>
|
| 55 |
|
| 56 |
**Figure 2:** Subsampling and Ablation Strategy
|
| 57 |
|