leo-pekelis-gradient
commited on
Commit
•
5861db3
1
Parent(s):
698e052
Update README.md
Browse files
README.md
CHANGED
@@ -9,7 +9,7 @@ license: llama3
|
|
9 |
---
|
10 |
<img src="https://cdn-uploads.huggingface.co/production/uploads/655bb613e8a8971e89944f3e/TSa3V8YpoVagnTYgxiLaO.png" width="200"/>
|
11 |
|
12 |
-
# Llama-3 8B Instruct 1048k
|
13 |
Gradient incorporates your data to deploy autonomous assistants that power critical operations across your business. To learn more or collaborate on a custom model, drop us a message at contact@gradient.ai.
|
14 |
|
15 |
This model extends LLama-3 8B's context length from 8k to > 1040K, developed by Gradient, sponsored by compute from [Crusoe Energy](https://huggingface.co/crusoeai). It demonstrates that SOTA LLMs can learn to operate on long context with minimal training by appropriately adjusting RoPE theta. We trained on 320M total tokens, which is < 0.002% of Lamma-3's original pre-training data.
|
@@ -39,13 +39,13 @@ For training data, we generate long contexts by augmenting [SlimPajama](https://
|
|
39 |
| Initialize From | LLaMA-3 7B| 65K | 262K | 524k |
|
40 |
| Sequence Length 2^N | 16 | 18 | 19 | 20 |
|
41 |
| RoPE theta | 15.3 M | 207.1 M | 1.06B | 2.80B |
|
42 |
-
|
|
43 |
-
|
|
44 |
| Steps | 30 | 24 | 50 | 50 |
|
45 |
| Total Tokens | 62914560 | 100663296 | 419430400 | 838860800 |
|
46 |
-
|
|
47 |
| # GPUs | 8 | 32 | 512 | 512 |
|
48 |
-
| Ring
|
49 |
| GPU Type | NVIDIA L40S | NVIDIA L40S | NVIDIA L40S | NVIDIA L40S |
|
50 |
| Minutes to Train (Wall)| 202 | 555 | 61 | 87 |
|
51 |
|
|
|
9 |
---
|
10 |
<img src="https://cdn-uploads.huggingface.co/production/uploads/655bb613e8a8971e89944f3e/TSa3V8YpoVagnTYgxiLaO.png" width="200"/>
|
11 |
|
12 |
+
# Llama-3 8B Gradient Instruct 1048k
|
13 |
Gradient incorporates your data to deploy autonomous assistants that power critical operations across your business. To learn more or collaborate on a custom model, drop us a message at contact@gradient.ai.
|
14 |
|
15 |
This model extends LLama-3 8B's context length from 8k to > 1040K, developed by Gradient, sponsored by compute from [Crusoe Energy](https://huggingface.co/crusoeai). It demonstrates that SOTA LLMs can learn to operate on long context with minimal training by appropriately adjusting RoPE theta. We trained on 320M total tokens, which is < 0.002% of Lamma-3's original pre-training data.
|
|
|
39 |
| Initialize From | LLaMA-3 7B| 65K | 262K | 524k |
|
40 |
| Sequence Length 2^N | 16 | 18 | 19 | 20 |
|
41 |
| RoPE theta | 15.3 M | 207.1 M | 1.06B | 2.80B |
|
42 |
+
| Batch Size | 1 | 1 | 2 | 2 |
|
43 |
+
| Gradient Accumulation Steps | 32 | 16 | 1 | 1 |
|
44 |
| Steps | 30 | 24 | 50 | 50 |
|
45 |
| Total Tokens | 62914560 | 100663296 | 419430400 | 838860800 |
|
46 |
+
| Learning Rate | 2.00E-05 | 2.00E-05 | 2.00E-05 | 2.00E-05 |
|
47 |
| # GPUs | 8 | 32 | 512 | 512 |
|
48 |
+
| Ring parallelism | 1 | 1 | 8 | 8 |
|
49 |
| GPU Type | NVIDIA L40S | NVIDIA L40S | NVIDIA L40S | NVIDIA L40S |
|
50 |
| Minutes to Train (Wall)| 202 | 555 | 61 | 87 |
|
51 |
|