[WIP] Fix float32 precision (multi-commit 6f4f3fee98ff848b483b84b692a4b03251126f470852d10f1205f6af709627e7)
#5
by
keyfan
- opened
README.md
CHANGED
@@ -5,11 +5,4 @@ library_name: transformers
|
|
5 |
|
6 |
Unofficial dequantized weight of [grok-1](https://huggingface.co/xai-org/grok-1) in HF Transformers format.
|
7 |
|
8 |
-
|
9 |
-
|
10 |
-
The (fp32) weights are converted using the [script here](https://gist.github.com/chu-tianxiang/ec310e15d56949fd0f351cb5f65ee7a1) ran inside the [grok-1 repo](https://github.com/xai-org/grok-1). Since downloading the dequantized weight needs twice as much time, it's recommended to download the original weight and convert on your own.
|
11 |
-
|
12 |
-
#### Benchmarks
|
13 |
-
(I ran with `load_in_8bit` using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) due to limited hardware, so the result will be slightly worse)
|
14 |
-
* MMLU 5-shot: 0.7166
|
15 |
-
* BBH 3-shot: 0.5204
|
|
|
5 |
|
6 |
Unofficial dequantized weight of [grok-1](https://huggingface.co/xai-org/grok-1) in HF Transformers format.
|
7 |
|
8 |
+
The weights are converted using the [script here](https://gist.github.com/chu-tianxiang/ec310e15d56949fd0f351cb5f65ee7a1) ran inside the [grok-1 repo](https://github.com/xai-org/grok-1). Since downloading the dequantized weight needs twice as much time, it's recommended to download the original weight and convert on your own.
|
|
|
|
|
|
|
|
|
|
|
|
|
|