vadimbelsky/Qwen3-4B-Instruct-2507-MLX-4bit-Vadim-Belski

Model overview

This repository contains a LoRA adapter trained to infuse a base language model with domain knowledge, persona, and technical expertise based on the public CV and authored content of Vadim (Vadzim) Belski. The adapter is intended to improve the base model's ability to answer questions about Vadim's background, projects, architecture practices, AI/ML approach, blockchain experience, and related enterprise engineering topics.

Training data

The adapter was trained on a small curated dataset derived from Vadim Belski's CV, website and provided project descriptions. Data format: newline-delimited JSON (data/train.jsonl). Each line is a single JSON object with a text field containing a user prompt and an assistant response.

Key characteristics:

• ~534 records after synthetic augmentation (original CV-derived lines + synthetic Q/A examples).
• Content focuses on: career history, project summaries (Healthcare Voice AI, Wine Digital Passport, Secure Wallet, etc.), technology stack (Python, Golang, AWS, Kubernetes), AI practices (RAG, agents), and security/compliance practices.
• Synthetic augmentation: programmatically generated Q/A pairs to increase coverage of engineering topics and best practices.

Provenance and licenses:

• Primary source: user-provided CV and website content (https://belski.me). The user attests they have rights to use this content for fine-tuning.
• Synthetic content was generated by scripts in scripts/ and derived from the original CV facts.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("vadimbelsky/Qwen3-4B-Instruct-2507-MLX-4bit-Vadim-Belski")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)