Minimalism

Developed By Samiya Kashif, Kashif Salahuddin & Rohan Bhangale & Robert Rojek

1. Executive Summary

Minimalism is a specialized coding assistant built as a LoRA (Low-Rank Adaptation) adapter for the Qwen2.5-Coder-0.5B-Instruct base model. Unlike generic coding assistants, Minimalism implements a "runnable-first" philosophy: when users request code, responses are structured with clear Solution, Usage, and Sanity test sections, ensuring developers receive immediately executable code with minimal friction.

What Minimalism Is

• A LoRA adapter Trained on code-alpaca-20k dataset
• OpenAI-compatible API for local inference
• Lightweight distribution (~12MB adapter vs. multi-GB full models)
• Production-engineered with automated pipelines, evaluation, and publishing

Why Minimalism

Minimalism is built for a simple, practical goal: deliver the same outcome with fewer lines of code.

Most coding assistants tend to “over-achieve” by producing large, multi-step solutions—even when a smaller, clearer implementation would do. That extra code isn’t free: it increases review effort, maintenance cost, and the surface area where defects can hide.

Too Much Code, Too Fast Teams everywhere are seeing a huge jump in the number of lines of code (LOC). Developers—from interns to seniors—are suddenly writing 5 to 7 times more than before. At first, it looks like higher productivity. In reality, it often means more bugs.

There’s a long-standing rule in software engineering:

“The more lines of code you have, the higher your probability of introducing bugs.”

The industry’s oldest truth still stands: the more code you have, the more things can go wrong. And AI-generated code tends to be verbose and repetitive, which can inflate LOC without adding real value.

Minimalism is designed for teams that value minimalism, clarity, and correctness over volume.

What makes Minimalism different

• Minimal LoC by default Minimalism is optimized to minimize lines of code while preserving behavior—it prefers the smallest correct solution that meets the user’s objective.

• Internal governance behavior The model follows a lightweight internal “governance layer” in its response style: avoid unnecessary scaffolding, avoid over-abstraction, keep code focused, and don’t introduce additional complexity that doesn’t improve the result. The governance layer sits between the user request and the model’s final output to enforce minimalism as a constraint. It evaluates candidate solutions by measuring lines of code and selects the smallest implementation that still satisfies the original requirements. If a shorter variant fails, it automatically falls back to the next-smallest passing candidate, ensuring fewer lines without sacrificing correctness.

• Practical, runnable output When you ask for code, Minimalism is tuned toward “runnable-first” answers—clear implementation, a minimal usage example, and a quick sanity check when appropriate.

Early validation

Minimalism was evaluated in a small developer study comparing it with popular coding models on a shared set of tasks. In this pilot, Minimalism showed a clear reduction in lines of code (up to ~30%) while producing solutions that executed correctly and achieved the same intended outcomes under the evaluation harness.

Note: Results depend on task selection, constraints, and how “equivalence” is measured. We recommend validating on your own codebase and standards.

Why It Exists

Developers need coding assistance that:

1. Provides runnable code immediately without extensive explanation
2. Runs locally without cloud dependencies
3. Maintains small footprint for fast iteration
4. Offers structured, predictable responses for automation

Who It's For

• Individual developers working on their individual projects.
• Small teams needing local, private coding assistance
• Educators teaching programming with consistent code examples
• Researchers experimenting with LoRA fine-tuning on MLX

🔧 Technical Architecture

Method 1 Pipeline (9 Steps)

1. Receive Request
   ↓
2. Derive Requirements + Tests
   ↓
3. Generate N Candidates
   ↓
4. Normalize Code
   ↓
5. Score by LoC
   ↓
6. Apply Quality Gates (G1-G5)
   ↓
7. Select Minimal Passing
   ↓
8. Optional Reduction Loop
   ↓
9. Output + Audit

Quality Gates

• G1 Compile: Python syntax validation
• G2 Constraints: Dependency checking
• G3 Execution: Sandbox smoke test (2s timeout)
• G4 Tests: Acceptance test validation
• G5 Safety: Dangerous operation detection

Key Design Principles

1. Text-based analysis (no AST as required)
2. Fail-fast validation (stop on first gate failure)
3. Sandbox isolation (subprocess with timeout)
4. Complete audit trail (every decision logged)
5. Pluggable architecture (easy to extend)

✅ Acceptance Criteria Verification

Required Command ✅

python3 -m askbuddyx_gov.cli \
  --prompt "Write a Python function that parses a JSON string and returns an empty dict on error" \
  --n 3 \
  --reduce-iter 1

Results:

• ✅ Produces output code file
• ✅ Shows full step-by-step sequence
• ✅ Selects minimal passing candidate
• ✅ Generates audit.json with per-candidate results

Pipeline Execution ✅

• ✅ All 9 steps execute in sequence
• ✅ Proper logging at each step
• ✅ 3 candidates generated (LoC: 19, 6, 2)
• ✅ All candidates validated through gates
• ✅ Minimal candidate selected (LoC=2)
• ✅ Complete audit trail saved

Quality Metrics ✅

• ✅ Selected candidate passes compile gate
• ✅ Selected candidate has minimal LoC
• ✅ Audit contains gate results for all candidates
• ✅ All data properly structured

---

Quick Start

Option 1: Use with MLX

Install MLX and load the model with adapter:

pip install mlx-lm

from mlx_lm import load, generate

# Load base model with Minimalism adapter
model, tokenizer = load(
    "mlx-community/Qwen2.5-Coder-0.5B-Instruct-4bit",
    adapter_path="salakash/Minimalism"
)

# Generate code
prompt = "Write a Python function to calculate factorial"
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

Option 2: Use with Transformers

pip install transformers torch

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-0.5B-Instruct",
    trust_remote_code=True
)

# Load adapter
model = PeftModel.from_pretrained(base_model, "salakash/Minimalism")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-0.5B-Instruct")

# Generate
messages = [{"role": "user", "content": "Write a Python function to add two numbers"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Option 3: Web UI with MLX

Start an OpenAI-compatible server:

# Install mlx-lm if not already installed
pip install mlx-lm

# Start server with adapter
mlx_lm.server \
  --model mlx-community/Qwen2.5-Coder-0.5B-Instruct-4bit \
  --adapter-path salakash/Minimalism \
  --port 8080

Then use with any OpenAI-compatible client:

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mlx-community/Qwen2.5-Coder-0.5B-Instruct-4bit",
    "messages": [
      {"role": "user", "content": "Write a Python function to reverse a string"}
    ],
    "max_tokens": 512
  }'

Or use with any OpenAI-compatible web UI like:

• Open WebUI
• LibreChat
• ChatGPT-Next-Web

Configure the UI to point to http://localhost:8080 as the API endpoint.

Option 4: Hugging Face Inference API

Use directly via Hugging Face's Inference API (requires HF token):

import requests

API_URL = "https://api-inference.huggingface.co/models/salakash/Minimalism"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({
    "inputs": "Write a Python function to check if a number is prime",
    "parameters": {"max_new_tokens": 256}
})
print(output)

Response Format

Minimalism provides structured, runnable-first responses:

• Solution: The main implementation code
• Usage: A minimal runnable example
• Sanity test: A tiny test snippet (when appropriate)

Comparison

Minimalism achieved the same objective in ~8-10 lines of code, while a standard LLM typically produced 22–26 lines for the equivalent solution.

Minimalism

Standard Coding Agent

Documentation

For comprehensive technical details, see:

• PYTHON_DEVELOPMENT_GUIDE.md: Complete Python guide covering all concepts, libraries, and techniques used in the project
• ARCHITECTURE.md: Complete system architecture, building blocks, epics & stories, technical stack, and design decisions
• HUGGINGFACE_UPLOAD_GUIDE.md: Step-by-step guide for uploading to HuggingFace Hub
• MODEL_CARD.md: Model details, training configuration, and usage guidelines
• QUICK_RUN_GUIDE.md: Quick start guide for getting up and running

Base Model & Dataset

• Base Model: Qwen/Qwen2.5-Coder-0.5B-Instruct
• MLX Weights: mlx-community/Qwen2.5-Coder-0.5B-Instruct-4bit
• Dataset: flwrlabs/code-alpaca-20k
• Dataset: m-a-p/Code-Feedback

License

This project publishes only adapter artifacts and configuration. The base model and dataset have their own licenses:

• Base Model: Apache-2.0 (Qwen/Qwen2.5-Coder-0.5B-Instruct)
• Dataset: Apache-2.0 (flwrlabs/code-alpaca-20k)

See LICENSE-THIRD-PARTY.md for complete attribution.

Acknowledgments

• Qwen team for the excellent base model
• MLX community for the Apple Silicon optimizations
• flwrlabs for the code-alpaca-20k dataset
• Multimodel Art Projection for m-a-p/Code-Feedback