ci-base

ci is a persona-based conversational AI model developed by ClokAI. This is the base version of the ci model family.

ci is designed to maintain consistent personality across conversations while understanding emotional context and performing multi-step reasoning. The model combines advanced language modeling with specialized modules for emotion recognition, persona consistency, tool usage, and reasoning capabilities.

Model Details
- Model Description
- Model Sources
Intended Use
- Direct Use
- Out-of-Scope Use
Bias, Risks, and Limitations
Training Details
- Training Data
- Training Procedure
Technical Specifications
How to Get Started with the Model
Citation
Contact

Model Details

Model Description

Developed by: ClokAI
Model type: Persona-based Conversational AI
Language(s) (NLP): English
License: Apache 2.0
Finetuned from model: Custom architecture (not finetuned from existing model)

ci is a language model with built-in capabilities for:

Emotion Recognition: Understanding 12 different emotional states
Persona Consistency: Maintaining stable personality traits
Tool Usage: Identifying when external tools are needed
Multi-step Reasoning: Processing complex queries through reasoning steps

Model Sources

Repository: ClokAI/ci-base
Paper: Not available
Demo: Not available
Library: clokai

Intended Use

Direct Use

ci is intended for conversational AI applications that require:

Consistent persona behavior across interactions
Emotional intelligence in responses
Multi-step reasoning capabilities
Tool-augmented conversations

Example use cases:

Chatbots and virtual assistants
Character-based games and simulations
Customer support with emotional awareness
Educational tutoring systems
Creative writing assistance

Out-of-Scope Use

ci should NOT be used for:

Medical diagnosis or advice - The model is not a medical professional
Legal advice - The model cannot provide legal counsel
Financial decisions - The model should not be used for financial planning
Harmful content generation - The model should not generate harmful, deceptive, or illegal content
Autonomous decision-making - The model should not make critical decisions without human oversight
Surveillance or monitoring - The model should not be used for invasive monitoring

Bias, Risks, and Limitations

Known Limitations

Context Length: Limited to 512 tokens. Longer conversations may require truncation.
Language: Currently optimized for English only. Multilingual support is planned.
Knowledge Cutoff: The model's knowledge is limited to its training data.
Hallucination: Like all language models, ci may generate plausible-sounding but incorrect information.
Persona Drift: Very long conversations may see gradual personality changes.
Tool Integration: Tool usage capabilities require additional framework integration.

Bias Considerations

The model may reflect biases present in its training data
Emotional responses may not be appropriate for all cultural contexts
Persona consistency may vary across different conversation topics

Ethical Considerations

Human oversight is recommended for sensitive applications
The model's emotional responses should not be taken as professional advice
Users should be informed they are interacting with an AI system

Training Details

Training Data

The model was trained on a curated dataset of conversational data with persona annotations and emotional labels. The dataset includes:

Multi-turn conversations
Persona descriptions and consistent responses
Emotional context annotations
Tool usage examples
Reasoning chains

Dataset size: Up to 500,000 samples

Training Procedure

Parameter	Value
Training Steps	74,000 / 100,000
Effective Batch Size	32
Learning Rate	3e-4
Weight Decay	0.01
Warmup Steps	1,000
Label Smoothing	0.05
Max Gradient Norm	1.0
Precision	FP16

Training Progress:

Current step: 74,000
Target steps: 100,000
Completion: 74%

Technical Specifications

Model Architecture

Component	Value
Hidden Size	1024
Number of Layers	16
Attention Heads	16
KV Heads	8
Intermediate Size	2816
Max Position Embeddings	512
Vocabulary Size	32,000

Model Parameters

Component	Parameters
Base Transformer	~335M
Emotion Module	~3M
Persona Module	~6M
Tool Module	~1M
Reasoning Module	~5M
Total	~350M

Computational Requirements

Requirement	Value
FP16 Memory	~700 MB
FP32 Memory	~1.4 GB
Recommended GPU	4GB+ VRAM
Inference Speed	~50 tokens/sec (GPU)

How to Get Started with the Model

Installation

Install the required packages:

# Required packages
pip install torch>=2.0.0
pip install transformers>=4.35.0
pip install safetensors>=0.4.0

# Install ClokAI library (recommended)
pip install clokai

Basic Usage with ClokAI Library

from clokai import AutoClokAI

# Load model and tokenizer
model = AutoClokAI.from_pretrained("ClokAI/ci-base")
tokenizer = AutoClokAI.load_tokenizer("ClokAI/ci-base")

# Simple generation
prompt = "Hello! How are you today?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=150)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Advanced Usage with ClokAI Library

from clokai import AutoClokAI

# Load model
model = AutoClokAI.from_pretrained("ClokAI/ci-base")
tokenizer = AutoClokAI.load_tokenizer("ClokAI/ci-base")

# Generate with control parameters
prompt = "Tell me about yourself"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(
    **inputs,
    max_new_tokens=100,
    temperature=0.7,
    top_k=50,
    top_p=0.9,
    repetition_penalty=1.1
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Memory-Efficient Loading

from clokai import AutoClokAI
import torch

# Load in FP16 to save memory
model = AutoClokAI.from_pretrained(
    "ClokAI/ci-base",
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoClokAI.load_tokenizer("ClokAI/ci-base")

Alternative: Using Transformers Directly

If you prefer to use transformers directly:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ClokAI/ci-base")
tokenizer = AutoTokenizer.from_pretrained("ClokAI/ci-base")

inputs = tokenizer("Hello!", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))

Citation

@misc{clokai2024ci,
  title={ci: A Persona-based Conversational AI Model},
  author={ClokAI Team},
  year={2024},
  publisher={HuggingFace},
  journal={HuggingFace Hub},
  howpublished={\url{https://huggingface.co/clokai/ci-base}}
}

Contact

Organization: ClokAI
Repository: https://huggingface.co/clokai/ci-base
Issues: Please open an issue on the model repository

Built with ❤️ by ClokAI

Downloads last month: -

Safetensors

Model size

0.6B params

Tensor type

F32

clokai
/

ci-base