Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: upstage/Solar-Open-100B
|
| 4 |
+
tags:
|
| 5 |
+
- pruned
|
| 6 |
+
- compressed
|
| 7 |
+
- llm
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# Solar-Open-100B-pruned-5pct
|
| 11 |
+
|
| 12 |
+
This model is a pruned version of [upstage/Solar-Open-100B](https://huggingface.co/upstage/Solar-Open-100B).
|
| 13 |
+
|
| 14 |
+
## Pruning Details
|
| 15 |
+
|
| 16 |
+
| Property | Value |
|
| 17 |
+
|----------|-------|
|
| 18 |
+
| Original Model | upstage/Solar-Open-100B |
|
| 19 |
+
| Original Parameters | 43.71B |
|
| 20 |
+
| Pruned Parameters | 7.21B |
|
| 21 |
+
| Compression Ratio | 0.1650 (6.1x smaller) |
|
| 22 |
+
| Strategy | Layer + FFN (Recommended) |
|
| 23 |
+
| Importance Metric | magnitude |
|
| 24 |
+
| Layers Removed | 28 |
|
| 25 |
+
| FFN Reduction | 0.00% |
|
| 26 |
+
|
| 27 |
+
## Usage
|
| 28 |
+
|
| 29 |
+
```python
|
| 30 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 31 |
+
|
| 32 |
+
model = AutoModelForCausalLM.from_pretrained("llaa33219/Solar-Open-100B-pruned-5pct")
|
| 33 |
+
tokenizer = AutoTokenizer.from_pretrained("llaa33219/Solar-Open-100B-pruned-5pct")
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
## Notes
|
| 37 |
+
|
| 38 |
+
This model was created using structured pruning techniques including:
|
| 39 |
+
- Layer pruning (removing entire transformer layers)
|
| 40 |
+
- FFN dimension pruning (reducing intermediate layer sizes)
|
| 41 |
+
|
| 42 |
+
The pruning was based on magnitude importance scoring to preserve the most important weights.
|