Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,144 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
---
|
| 4 |
+
Model Summary
|
| 5 |
+
|
| 6 |
+
OpenCelestial_1 is a compact and efficient language model fine-tuned on a greeting dataset. It demonstrates that small LLMs can achieve remarkable conversational capabilities, even when trained on consumer-grade hardware.
|
| 7 |
+
|
| 8 |
+
Based on the GPT-2 architecture, OpenCelestial_1 is optimized for clear, polite, and structured responses, making it ideal for use cases such as:
|
| 9 |
+
|
| 10 |
+
Chatbots
|
| 11 |
+
Instruction-following assistants
|
| 12 |
+
Lightweight deployments on limited hardware
|
| 13 |
+
|
| 14 |
+
Model Training
|
| 15 |
+
|
| 16 |
+
Base Model: openai-community/gpt2
|
| 17 |
+
Dataset: Custom greeting dataset with structured "User" and "AI" dialogue pairs.
|
| 18 |
+
Hardware: Fine-tuned on a single NVIDIA RTX 3060.
|
| 19 |
+
Optimization: Fine-tuning utilized LoRA (Low-Rank Adaptation) to improve memory efficiency.
|
| 20 |
+
|
| 21 |
+
Usage Example
|
| 22 |
+
|
| 23 |
+
To interact with OpenCelestial_1, use the following Python script:
|
| 24 |
+
|
| 25 |
+
pip install transformers torch
|
| 26 |
+
|
| 27 |
+
Copy and paste the following script:
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
```python3
|
| 32 |
+
from transformers import GPT2LMHeadModel, GPT2Tokenizer
|
| 33 |
+
import torch
|
| 34 |
+
|
| 35 |
+
# Load the model and tokenizer
|
| 36 |
+
model_path = "./gpt2_lora_alpaca_gpt4"
|
| 37 |
+
model = GPT2LMHeadModel.from_pretrained(model_path)
|
| 38 |
+
tokenizer = GPT2Tokenizer.from_pretrained(model_path)
|
| 39 |
+
|
| 40 |
+
# Set the pad token to the EOS token if not already set
|
| 41 |
+
tokenizer.pad_token = tokenizer.eos_token
|
| 42 |
+
|
| 43 |
+
print("Chatbot is ready! Type 'exit' to quit.")
|
| 44 |
+
|
| 45 |
+
while True:
|
| 46 |
+
user_input = input("You: ")
|
| 47 |
+
if user_input.lower() == "exit":
|
| 48 |
+
print("Chatbot: Goodbye!")
|
| 49 |
+
break
|
| 50 |
+
|
| 51 |
+
# Define the system prompt and the full prompt
|
| 52 |
+
system_prompt = "You are an intelligent AI assistant that will answer every question to the best of your ability. Be clear and polite with your answers."
|
| 53 |
+
prompt = f"{system_prompt}\n### Instruction:\n{user_input}\n### Response:"
|
| 54 |
+
|
| 55 |
+
# Tokenize the input
|
| 56 |
+
inputs = tokenizer(
|
| 57 |
+
prompt,
|
| 58 |
+
return_tensors="pt",
|
| 59 |
+
padding=True,
|
| 60 |
+
truncation=True,
|
| 61 |
+
max_length=1024,
|
| 62 |
+
)
|
| 63 |
+
input_ids = inputs.input_ids.to(model.device)
|
| 64 |
+
attention_mask = inputs.attention_mask.to(model.device)
|
| 65 |
+
|
| 66 |
+
# Generate the response
|
| 67 |
+
with torch.no_grad():
|
| 68 |
+
outputs = model.generate(
|
| 69 |
+
input_ids=input_ids,
|
| 70 |
+
attention_mask=attention_mask,
|
| 71 |
+
max_new_tokens=150,
|
| 72 |
+
pad_token_id=tokenizer.eos_token_id,
|
| 73 |
+
do_sample=True,
|
| 74 |
+
temperature=0.7,
|
| 75 |
+
top_k=50,
|
| 76 |
+
top_p=0.95,
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
# Decode the response and clean it up
|
| 80 |
+
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 81 |
+
clean_response = response.split("### Response:")[-1].strip()
|
| 82 |
+
print(f"Chatbot: {clean_response}")
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
Example Outputs
|
| 86 |
+
|
| 87 |
+
Prompt: Hello there!
|
| 88 |
+
Response: Hello there! I am just an AI assistant, but I’m here to help you with anything you need.
|
| 89 |
+
|
| 90 |
+
Prompt: Can you tell me a joke?
|
| 91 |
+
Response: Sure! Why don’t skeletons fight each other? Because they don’t have the guts!
|
| 92 |
+
|
| 93 |
+
Prompt: What is the capital of France?
|
| 94 |
+
Response: The capital of France is Paris.
|
| 95 |
+
Training Details
|
| 96 |
+
|
| 97 |
+
LoRA Configuration:
|
| 98 |
+
Rank (r): 4
|
| 99 |
+
Alpha: 16
|
| 100 |
+
Dropout: 0.1
|
| 101 |
+
Target Modules: GPT-2’s attention layers (attn.c_attn)
|
| 102 |
+
Training Arguments:
|
| 103 |
+
Mixed precision: Enabled (fp16)
|
| 104 |
+
Epochs: 3
|
| 105 |
+
Batch size: 2 (to fit GPU memory)
|
| 106 |
+
Learning rate: 5e-5
|
| 107 |
+
|
| 108 |
+
Performance
|
| 109 |
+
|
| 110 |
+
OpenCelestial_1 demonstrates:
|
| 111 |
+
|
| 112 |
+
Clear conversational ability with polite, structured responses.
|
| 113 |
+
Low resource requirements, suitable for GPUs like the RTX 3060.
|
| 114 |
+
Consistency in instruction-following tasks.
|
| 115 |
+
|
| 116 |
+
Intended Use
|
| 117 |
+
|
| 118 |
+
This model is designed for:
|
| 119 |
+
|
| 120 |
+
Conversational AI applications.
|
| 121 |
+
Instruction-based assistants that respond politely and clearly.
|
| 122 |
+
Lightweight deployments for hobbyists, small-scale developers, or educational purposes.
|
| 123 |
+
|
| 124 |
+
Limitations
|
| 125 |
+
|
| 126 |
+
Responses may still contain hallucinations or factual inaccuracies.
|
| 127 |
+
Performance is limited to the dataset scope and GPT-2’s inherent capabilities.
|
| 128 |
+
|
| 129 |
+
Citation
|
| 130 |
+
|
| 131 |
+
If you use OpenCelestial_1 in your work, please consider citing:
|
| 132 |
+
|
| 133 |
+
@misc{OpenCelestial_1,
|
| 134 |
+
author = {Your Name or Organization},
|
| 135 |
+
title = {OpenCelestial_1: A Compact GPT-2 Fine-Tuned Model},
|
| 136 |
+
year = {2024},
|
| 137 |
+
howpublished = {\url{https://huggingface.co/your_username/OpenCelestial_1}},
|
| 138 |
+
}
|
| 139 |
+
|
| 140 |
+
Acknowledgments
|
| 141 |
+
|
| 142 |
+
Base Model: openai-community/gpt2
|
| 143 |
+
Fine-tuned using the LoRA technique for efficient memory usage.
|
| 144 |
+
Developed on a single NVIDIA RTX 3060 GPU.
|