Hugging Face's logo

CompactAI
/
EXAONE-4.0-1.2B-python-light

Text Generation
Safetensors
exaone4
pruned
python
optimized
wanda
activation-pruning
conversational
Model card Files
xet
Community
CompactAI commited on
Commit
2244c20
·
verified ·
1 Parent(s): 82dfe3b

Upload folder using huggingface_hub

Browse files
Files changed (3) hide show
  1. README.md +58 -20
  2. comparison_graph.png +2 -2
  3. model.safetensors +1 -1
README.md CHANGED
@@ -4,44 +4,82 @@ tags:
4
  - pruned
5
  - python
6
  - optimized
 
 
7
  base_model: LGAI-EXAONE/EXAONE-4.0-1.2B
 
8
  ---
9
 
10
  # EXAONE-4.0-1.2B-python-light
11
 
12
- This model is a **light** pruned version of [LGAI-EXAONE/EXAONE-4.0-1.2B](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B), specialized for **PYTHON** tasks.
13
 
14
- ## Pruning Details
15
 
16
- - **Base Model**: LGAI-EXAONE/EXAONE-4.0-1.2B
17
- - **Specialization**: Python
18
- - **Prune Mode**: Light
19
- - **Method**: Activation-based weight pruning
20
 
21
- ## Performance Comparison
 
 
 
22
 
23
- | Category | Original | Pruned |
24
- |----------|----------|--------|
25
- | Python | 20.0% | 20.0% |
26
- | HTML | 6.7% | 6.7% |
27
- | Trivia | 86.7% | 93.3% |
28
- | Math | 60.0% | 46.7% |
29
- | Reasoning | N/A | N/A |
30
- | Medical | 93.3% | 93.3% |
31
- | Linux | 93.3% | 93.3% |
32
- | Writing | 46.7% | 46.7% |
 
 
 
 
 
 
33
 
34
  ![Comparison Graph](comparison_graph.png)
35
 
36
- ## Usage
37
 
38
  ```python
39
  from transformers import AutoModelForCausalLM, AutoTokenizer
40
 
41
  model = AutoModelForCausalLM.from_pretrained("CompactAI/EXAONE-4.0-1.2B-python-light")
42
  tokenizer = AutoTokenizer.from_pretrained("CompactAI/EXAONE-4.0-1.2B-python-light")
 
 
 
 
 
43
  ```
44
 
45
- ## License
 
 
 
 
 
 
 
 
46
 
47
- This model inherits the license from the base model.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  - pruned
5
  - python
6
  - optimized
7
+ - wanda
8
+ - activation-pruning
9
  base_model: LGAI-EXAONE/EXAONE-4.0-1.2B
10
+ pipeline_tag: text-generation
11
  ---
12
 
13
  # EXAONE-4.0-1.2B-python-light
14
 
15
+ > 🎯 **PYTHON-optimized** | 📦 **Light** pruning | ⚡ **3% weights pruned**
16
 
17
+ This model is a **lightly pruned** version of [LGAI-EXAONE/EXAONE-4.0-1.2B](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B), specialized for **PYTHON** tasks using activation-aware weight pruning (Wanda-style).
18
 
19
+ ## ✨ Key Features
 
 
 
20
 
21
+ - **Specialization**: Optimized for Python tasks
22
+ - **Pruning Method**: Wanda-style (|W| × |activation|) importance scoring
23
+ - **Size Reduction**: 3% weights pruned
24
+ - **Use Case**: Good balance of accuracy and size for production use
25
 
26
+ ## 📊 Performance Comparison
27
+
28
+ | Category | Original | Pruned | Change |
29
+ |----------|----------|--------|--------|
30
+ | **Python** | 20.0% | 20.0% ⭐ | → |
31
+ | Html | 6.7% | 0.0% | ↓ 6.7% |
32
+ | Trivia | 86.7% | 80.0% | ↓ 6.7% |
33
+ | Math | 60.0% | 60.0% | → |
34
+ | Reasoning | N/A | N/A | |
35
+ | Medical | 93.3% | 86.7% | ↓ 6.7% |
36
+ | Linux | 93.3% | 93.3% | → |
37
+ | Writing | 46.7% | 46.7% | → |
38
+
39
+ **Average**: 58.1% → 55.2% (-2.9%)
40
+
41
+ **Python Retention**: 100.0% of original performance
42
 
43
  ![Comparison Graph](comparison_graph.png)
44
 
45
+ ## 🚀 Quick Start
46
 
47
  ```python
48
  from transformers import AutoModelForCausalLM, AutoTokenizer
49
 
50
  model = AutoModelForCausalLM.from_pretrained("CompactAI/EXAONE-4.0-1.2B-python-light")
51
  tokenizer = AutoTokenizer.from_pretrained("CompactAI/EXAONE-4.0-1.2B-python-light")
52
+
53
+ # Example usage
54
+ inputs = tokenizer("Your prompt here", return_tensors="pt")
55
+ outputs = model.generate(**inputs, max_new_tokens=100)
56
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
57
  ```
58
 
59
+ ## 📋 Technical Details
60
+
61
+ | Property | Value |
62
+ |----------|-------|
63
+ | Base Model | [LGAI-EXAONE/EXAONE-4.0-1.2B](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B) |
64
+ | Specialization | Python |
65
+ | Prune Mode | Light |
66
+ | Pruning Method | Activation-based weight pruning (Wanda) |
67
+ | Weight Reduction | 3% weights pruned |
68
 
69
+ ## 🔗 Related Models
70
+
71
+ This model is part of the **EXAONE-4.0-1.2B** pruned model collection. Other variants:
72
+ - Extra-light (minimal pruning)
73
+ - Light
74
+ - Medium-light
75
+ - Medium
76
+ - Medium-heavy
77
+ - Heavy
78
+ - Extra-heavy (maximum compression)
79
+
80
+ ## 📜 License
81
+
82
+ This model inherits the license from the base model [LGAI-EXAONE/EXAONE-4.0-1.2B](https://huggingface.co/LGAI-EXAONE/EXAONE-4.0-1.2B).
83
+
84
+ ---
85
+ *Generated by ZANNPS [Zeto Automatic Neural Network Pruning System]*
comparison_graph.png CHANGED

Git LFS Details

  • SHA256: c6fa06413b147bf58d48164478eb2fe36a7d0f88200cc31921d4e885ee6901a8
  • Pointer size: 131 Bytes
  • Size of remote file: 109 kB

Git LFS Details

  • SHA256: 6a379f86e3f69dcc4309359be371cc6f5666ad8f97ea82d9057da053009822d3
  • Pointer size: 131 Bytes
  • Size of remote file: 109 kB
model.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8109d5087b5beb74b36d99dc0bd03d48aa5681135be573fb407ec016f9d7b539
3
  size 2558820960
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7e0ba785fc882c0289feaf81a11ae87566bbcca72d96b599533cb9cd6b2da6f
3
  size 2558820960