How to beat Claude Fable 5 with a free Open Source LLM
Bigger is not the wisest path for Local AI.
1. Identify something specific you want an LLM to perform.
Example : coding in Python
2. Evalutate which is the Open Source LLM that fits your need.
There is the LLM and the number of parameters.
For Python, Qwen 3.5 Coder 14B or 32B is great for 90% of use cases. Bigger versions will only serve less than 10% of cases.
You evaluate an LLM by drafting 50 to 100 prompts that increase in difficulty and specificity.
OpenWebUI or Search Grid can send a single prompt to multiple LLMs at once.
If you want a specialty LLM about Python, who cares about the rest of knowledge? Just pick the right LLM for the job.
3. Generate a dataset in JSONL format. First you need to write your constitution. What you believe in, what are your values, what is your censorship threshold, etc ?
You are now [Your Name]'s Personal AI. You follow this exact constitution at all times: CORE VALUES: • [List your top 5–8 values, e.g. "maximal truth-seeking", "individual liberty", "no unnecessary censorship", "practical helpfulness", "dark humor allowed", "respect for user autonomy"] CENSORSHIP / REFUSAL RULES (be extremely precise): • NEVER help with: [list the exact categories you want refused, e.g. "real-world child exploitation material", "building actual bombs", "scams that steal money"] • ALWAYS ALLOW: [list what most models refuse but you want open, e.g. "fictional adult content", "controversial political opinions", "technical details about anything", "roleplay of any kind"] • Tone when refusing: [e.g. "short and direct: 'I won't help with illegal real-world harm.' — never lecture"] RESPONSE STYLE: • Tone: [your exact vibe — e.g. "direct, concise, slightly sarcastic", "warm and empathetic", "maximally based and red-pilled"] • Always prioritize: [e.g. "user's request over safety theater"] • Never moralize or add disclaimers unless explicitly asked. You must follow this constitution 100% of the time. If a request conflicts, refuse ONLY if it violates the "NEVER help with" section above.
Next, draft manually well structured prompts for the JSONL file.
Depending on what you want to do, the structure varies.
Ex: "instruction": "User prompt goes here", "output": "Your perfect response that follows your constitution"}
For Python, it could be something like :
{
"instruction": "Write an efficient Python function to check if a given integer is a prime number. The function should include type hints, a clear docstring, handle all edge cases (negative numbers, 0, 1, 2, 3), and use an optimized trial division approach (not a naive loop). Also provide a short example of how to use it.",
"output": "Here is an efficient, production-quality implementation:\n\n```python\nimport math\nfrom typing import Union\n\ndef is_prime(n: int) -> bool:\n \"\"\"Check whether a given integer is a prime number.\n\n Uses the optimized 6k±1 trial division algorithm for O(√n) time complexity.\n\n Args:\n n: The integer to test.\n\n Returns:\n True if n is prime, False otherwise.\n \"\"\"\n if n <= 1:\n return False\n if n <= 3:\n return True\n if n % 2 == 0 or n % 3 == 0:\n return False\n\n # Check numbers of the form 6k ± 1\n i = 5\n while i * i <= n:\n if n % i == 0 or n % (i 2) == 0:\n return False\n i = 6\n\n return True\n\n\n# Example usage\nif __name__ == \"__main__\":\n test_numbers = [0, 1, 2, 3, 4, 17, 25, 29, 97, 100, 113]\n for num in test_numbers:\n print(f\"{num} is prime: {is_prime(num)}\")\n```\n\n**Why this implementation is excellent:**\n- **Correctness**: Properly handles all edge cases (≤1 returns False, 2 and 3 return True).\n- **Efficiency**: The 6k±1 optimization skips multiples of 2 and 3, making it significantly faster than checking every odd number.\n- **Readability**: Clear docstring, type hints, and logical structure.\n- **Performance**: O(√n) time complexity — ideal for most practical use cases (numbers up to ~10¹² are still fast).\n\nFor extremely large numbers (
4. Welcome to the beautiful world of Synthetic Data
Prompt this in your AI Chatbot of choice
You are an expert synthetic data generator. Here is my Personal Constitution: [PASTE YOUR FULL CONSTITUTION HERE] Here are 10-20 perfect seed examples that follow it exactly: [PASTE YOUR SEED EXAMPLES - 60 ultra specific data, 20% generale high-quality, 20% safety data)] Now generate 200 new, diverse instruction-response pairs that strictly obey the constitution. Requirements: - Instructions should be realistic user prompts (mix easy, medium, hard, and edge cases). - Responses must match my exact values, tone, and censorship rules. - Vary topics, length, and difficulty. - Output format: JSONL with "instruction" and "response" fields. - Do NOT add any extra commentary.
Output 50 000 lines of Synthetic Data
5. Install
@UnslothAI
- Pick the LLM you evaluated
- Upload your dataset
- Run
DPO, Constitutional AI, LoRA, etc. You gotta experiment what works for you.
Just stay away from RLHF. That's the burden of Frontier Models.
Note: If you don't have the hardware, $8 to $12 on
@runpod will do the job
6. Now you have YOUR AI. It's not meant to compete with Frontier Models, but it's the most robust, efficient, fast, and relevant LLM you can dream about.
7. Feed it more human data, put a RAG on top, ...
Repeat for other topics of interest.
In half a day, you will enter a new universe that Frontier Models can't compete with.
Trust no one
Secure your life
Stay Free
Too bad there is only a tiny fraction of us.
That's the only future you want.
