BerenMillidge
commited on
Commit
•
eec7dd4
1
Parent(s):
31be28f
Update README.md
Browse files
README.md
CHANGED
@@ -56,7 +56,19 @@ model = AutoModelForCausalLM.from_pretrained("Zyphra/Zamba2-7B", device_map="cud
|
|
56 |
|
57 |
Zamba2-7B-Instruct punches dramatically above its weight, achieving extremely strong instruction-following benchmark scores.
|
58 |
|
59 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
|
62 |
Moreover, due to its unique hybrid SSM architecture, Zamba2-7B-Instruct achieves extremely low inference latency and rapid generation with a significantly smaller memory footprint than comparable transformer-based models.
|
|
|
56 |
|
57 |
Zamba2-7B-Instruct punches dramatically above its weight, achieving extremely strong instruction-following benchmark scores.
|
58 |
|
59 |
+
<div style="width: 50%; margin: auto;">
|
60 |
+
|
61 |
+
| Task | Score |
|
62 |
+
|:------------:|:---------:|
|
63 |
+
| IFEval | 69.95 |
|
64 |
+
| BBH | 33.33 |
|
65 |
+
| MATH Lvl 5 | 13.57 |
|
66 |
+
| GPQA | 10.28 |
|
67 |
+
| MUSR | 8.21 |
|
68 |
+
| MMLU-PRO | 32.43 |
|
69 |
+
| **Average** | **27.96** |
|
70 |
+
|
71 |
+
</div>
|
72 |
|
73 |
|
74 |
Moreover, due to its unique hybrid SSM architecture, Zamba2-7B-Instruct achieves extremely low inference latency and rapid generation with a significantly smaller memory footprint than comparable transformer-based models.
|