Upload README.md with huggingface_hub
Browse files
README.md
ADDED
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# GPT-2 XL Compressed Model Weights
|
2 |
+
|
3 |
+
This dataset contains the compressed model weights from **tensor network compression** methodology applied to GPT-2 XL.
|
4 |
+
|
5 |
+
## π Files Included
|
6 |
+
|
7 |
+
### Compressed Model Weights (.pt files)
|
8 |
+
- `compressed_gpt2_xl_68.3%.pt` - Base compressed model (~68% compression)
|
9 |
+
- `compressed_gpt2_xl_68.3%_healed.pt` - Compressed + knowledge distillation healing
|
10 |
+
- `compressed_gpt2_xl_68.3%_enwik8_trained.pt` - Compressed + enwik8 fine-tuning
|
11 |
+
- `compressed_gpt2_xl_68.3%_enwik8_final.pt` - Final version after training
|
12 |
+
- `compressed_gpt2_xl_68.3%_enwik8_finetuned.pt` - Fine-tuned version
|
13 |
+
|
14 |
+
### Architecture & Metadata
|
15 |
+
- `model_architecture.pkl` - Compressed model architecture
|
16 |
+
- `*_metadata.json` - Training and compression metadata
|
17 |
+
|
18 |
+
## π¬ Methodology
|
19 |
+
|
20 |
+
Based on quantum-inspired tensor network compression:
|
21 |
+
- **Matrix Product Operator (MPO)** tensor network decomposition
|
22 |
+
- **68% parameter reduction** (1.56B β ~500M parameters)
|
23 |
+
- **Tensor network** compression technique
|
24 |
+
- **Knowledge distillation** healing process
|
25 |
+
|
26 |
+
## π Usage
|
27 |
+
|
28 |
+
```python
|
29 |
+
import torch
|
30 |
+
|
31 |
+
# Load compressed weights
|
32 |
+
model_weights = torch.load('compressed_gpt2_xl_68.3%_healed.pt', map_location='cpu')
|
33 |
+
|
34 |
+
# For ready-to-use model, see:
|
35 |
+
# https://huggingface.co/prompterminal/gpt2-compressed
|
36 |
+
```
|
37 |
+
|
38 |
+
## π Compression Stats
|
39 |
+
|
40 |
+
- **Original GPT-2 XL**: 1.56B parameters, ~6.2GB
|
41 |
+
- **Compressed Version**: ~500M parameters, ~1.98GB
|
42 |
+
- **Compression Ratio**: 68% reduction
|
43 |
+
- **Method**: MPO tensor networks + healing
|
44 |
+
|
45 |
+
## π― Files Recommended for Use
|
46 |
+
|
47 |
+
- **Best for inference**: `compressed_gpt2_xl_68.3%_healed.pt`
|
48 |
+
- **Best for fine-tuning**: `compressed_gpt2_xl_68.3%_enwik8_trained.pt`
|
49 |
+
- **Research/analysis**: All files + metadata
|
50 |
+
|
51 |
+
## π Citation
|
52 |
+
|
53 |
+
```bibtex
|
54 |
+
@misc{tensor_network_compression_2024,
|
55 |
+
title={GPT-2 XL Compressed using Tensor Network Methods},
|
56 |
+
author={prompterminal},
|
57 |
+
year={2024},
|
58 |
+
howpublished={HuggingFace Dataset}
|
59 |
+
}
|
60 |
+
```
|
61 |
+
|
62 |
+
## π Related
|
63 |
+
|
64 |
+
- **Ready-to-use model**: [prompterminal/gpt2-compressed](https://huggingface.co/prompterminal/gpt2-compressed)
|
65 |
+
- **Tensor network compression research**: Matrix Product Operator methods
|
66 |
+
|
67 |
+
---
|
68 |
+
*These weights represent pioneering work in tensor network compression for large language models.*
|