Spaces:
Sleeping
Sleeping
Update README.md
Browse files
README.md
CHANGED
@@ -12,3 +12,127 @@ short_description: Gradio
|
|
12 |
---
|
13 |
|
14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
---
|
13 |
|
14 |
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
15 |
+
|
16 |
+
GPT-2 Fine-Tuned TinyStories Project - FableWeaver AI
|
17 |
+
|
18 |
+
Overview
|
19 |
+
|
20 |
+
This project fine-tunes a GPT-2 model on the TinyStories dataset to generate structured, coherent, and engaging short narratives. The model is hosted on Hugging Face Spaces and provides a user-friendly interface for story generation.
|
21 |
+
|
22 |
+
Features
|
23 |
+
|
24 |
+
Story Generation: Produces coherent, child-friendly short stories.
|
25 |
+
|
26 |
+
Bias Monitoring: Ensures balanced gender and cultural representation.
|
27 |
+
|
28 |
+
Efficient Training: Fine-tuned on 200,000 training samples and 20,000 test samples.
|
29 |
+
|
30 |
+
Grammar & Readability Improvements: Integrated grammar-checking tools and text refinement.
|
31 |
+
|
32 |
+
Optimized Model Performance: Uses loss tracking, sampling techniques, and bias mitigation strategies.
|
33 |
+
|
34 |
+
System Architecture
|
35 |
+
|
36 |
+
The model is designed for easy interaction via Hugging Face Spaces and follows this pipeline:
|
37 |
+
|
38 |
+
Data Preprocessing & Cleaning
|
39 |
+
|
40 |
+
Tokenization, formatting, and encoding normalization.
|
41 |
+
|
42 |
+
Bias mitigation and balanced data preparation.
|
43 |
+
|
44 |
+
Model Fine-Tuning
|
45 |
+
|
46 |
+
Fine-tuned GPT-2 (124M parameters) using Hugging Face Transformers.
|
47 |
+
|
48 |
+
Hyperparameter optimization (batch size, learning rate, weight decay).
|
49 |
+
|
50 |
+
Story Generation Pipeline
|
51 |
+
|
52 |
+
Uses top-k filtering (k=50), top-p nucleus sampling (p=0.9), and temperature adjustments.
|
53 |
+
|
54 |
+
Post-Processing & Bias Mitigation
|
55 |
+
|
56 |
+
Contextual reinforcement and diversity-aware storytelling.
|
57 |
+
|
58 |
+
No-repeat n-gram settings and logical scene transitions.
|
59 |
+
|
60 |
+
Evaluation & Performance Monitoring
|
61 |
+
|
62 |
+
Tracked using Weights & Biases (W&B) and TensorBoard.
|
63 |
+
|
64 |
+
Analyzed validation loss and coherence checks.
|
65 |
+
|
66 |
+
Getting Started
|
67 |
+
|
68 |
+
Accessing the Model
|
69 |
+
|
70 |
+
The model is available on Hugging Face Spaces: GPT-2 TinyStories Generator
|
71 |
+
|
72 |
+
Usage Instructions
|
73 |
+
|
74 |
+
Visit the Hugging Face Space linked above.
|
75 |
+
|
76 |
+
Enter a prompt (e.g., "Once upon a time...") in the input field.
|
77 |
+
|
78 |
+
Click Generate to receive an AI-generated short story.
|
79 |
+
|
80 |
+
Modify the prompt and settings (temperature, top-k, top-p) for different results.
|
81 |
+
|
82 |
+
Training Details
|
83 |
+
|
84 |
+
Model: GPT-2 (124M)
|
85 |
+
|
86 |
+
Dataset: RonenEldan/TinyStories
|
87 |
+
|
88 |
+
Training: 3 epochs on Google Colab GPU (T4)
|
89 |
+
|
90 |
+
Loss Metrics:
|
91 |
+
|
92 |
+
Training Loss: 3.08 → 2.86
|
93 |
+
|
94 |
+
Validation Loss: 1.46 → 1.40
|
95 |
+
|
96 |
+
Evaluation & Observations
|
97 |
+
|
98 |
+
Perplexity improved from 8.12 → 2.09, indicating better text fluency.
|
99 |
+
|
100 |
+
Validation loss decreased consistently, suggesting effective generalization.
|
101 |
+
|
102 |
+
Human evaluation highlighted minor inconsistencies, such as abrupt scene shifts and simplistic narratives.
|
103 |
+
|
104 |
+
Ethical Considerations
|
105 |
+
|
106 |
+
Bias Monitoring: Pronoun analysis and diversity checks to ensure fairness.
|
107 |
+
|
108 |
+
Harmful Content Mitigation: Manually reviewed outputs for stereotypes.
|
109 |
+
|
110 |
+
Text Processing Issues: UTF-8 encoding applied to prevent character errors.
|
111 |
+
|
112 |
+
Future Improvements
|
113 |
+
|
114 |
+
Enhancing Creativity: Fine-tune temperature and randomness settings.
|
115 |
+
|
116 |
+
Genre-Specific Training: Introduce theme-based datasets.
|
117 |
+
|
118 |
+
Larger Model Training: Experiment with GPT-2 (355M) for richer storytelling.
|
119 |
+
|
120 |
+
Contributors
|
121 |
+
|
122 |
+
Charla Pia Vella (Project Developer)
|
123 |
+
|
124 |
+
Affiliation: ARI3333 Generative AI
|
125 |
+
|
126 |
+
License
|
127 |
+
|
128 |
+
This project is released under the Apache-2.0 License.
|
129 |
+
|
130 |
+
Acknowledgments
|
131 |
+
|
132 |
+
OpenAI for GPT-2
|
133 |
+
|
134 |
+
Hugging Face for the fine-tuning framework
|
135 |
+
|
136 |
+
Ronen Eldan for the TinyStories dataset
|
137 |
+
|
138 |
+
For more details, visit the Hugging Face Space.
|