WIDDX-AI-v1

WIDDX-AI-v1 is a fine-tuned conversational AI assistant based on Qwen2.5-7B-Instruct, optimized for instruction following and general conversation using LoRA (Low-Rank Adaptation) technique.

Model Details

Base Model: Qwen/Qwen2.5-7B-Instruct
Fine-tuning Method: LoRA (Low-Rank Adaptation)
Languages: English, Arabic
Model Type: Causal Language Model
License: Apache 2.0

Capabilities

This model excels at:

General Conversation: Natural, engaging dialogue
Instruction Following: Accurate execution of user instructions
Code Generation: Writing and explaining code in multiple programming languages
Question Answering: Providing informative responses to various topics
Text Analysis: Understanding and processing textual content
Creative Writing: Generating creative content and stories

Usage

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("WIDDX-AI/WIDDX-AI-v1", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("WIDDX-AI/WIDDX-AI-v1", trust_remote_code=True)

# Prepare input
messages = [
    {"role": "user", "content": "Explain machine learning in simple terms"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt")

# Generate response
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        temperature=0.7,
        top_p=0.9,
        do_sample=True
    )

response = tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True)
print(response)

Memory-Efficient Loading

For systems with limited GPU memory, use 4-bit quantization:

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch

# Configure quantization
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True
)

# Load with quantization
model = AutoModelForCausalLM.from_pretrained(
    "WIDDX-AI/WIDDX-AI-v1",
    quantization_config=quantization_config,
    device_map="auto",
    trust_remote_code=True
)

CPU Inference

For CPU-only inference:

model = AutoModelForCausalLM.from_pretrained(
    "WIDDX-AI/WIDDX-AI-v1",
    torch_dtype=torch.float32,
    trust_remote_code=True
)

Training Details

Dataset

Custom instruction-following dataset with high-quality examples
Diverse topics covering conversation, coding, analysis, and creativity
Carefully curated and filtered for quality and safety

Training Parameters

Method: LoRA (Low-Rank Adaptation)
Rank: 16
Alpha: 32
Dropout: 0.05
Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Learning Rate: 2e-4
Batch Size: 4 (with gradient accumulation)
Epochs: 3

Hardware

Training conducted on NVIDIA GPUs with 4-bit quantization
Compatible with both GPU and CPU inference

Performance

The model demonstrates strong performance across various tasks:

High instruction following accuracy
Natural conversational flow
Accurate code generation and explanation
Comprehensive knowledge across multiple domains

Limitations

May occasionally generate incorrect factual information
Performance depends on input quality and clarity
Limited by the knowledge cutoff of the base model
Requires appropriate prompt formatting for optimal results

Safety and Bias

This model inherits the safety measures and potential biases from its base model. Users should:

Review outputs for appropriateness
Implement additional safety measures in production
Be aware of potential cultural and linguistic biases

Citation

If you use this model in your research or applications, please cite:

@misc{widdx-ai-v1,
  title={WIDDX-AI-v1: A Fine-tuned Conversational Assistant},
  author={WIDDX-AI},
  year={2024},
  url={https://huggingface.co/WIDDX-AI/WIDDX-AI-v1}
}

Contact

For questions, issues, or collaboration opportunities, please visit our repository or contact the WIDDX-AI team.

Built with ❤️ by the WIDDX-AI team

WIDDX-AI
/

WIDDX-AI-v1