A Qwen2.5-Coder-3B fine-tuned on Cardano / Aiken smart-contract data, converted to Apple MLX 4-bit format for fast on-device inference on Apple Silicon.

Base model: mlx-community/Qwen2.5-Coder-3B-Instruct-4bit
Fine-tuning: LoRA (rank 8), 800 iterations on Cardano/Aiken code examples
Format: MLX 4-bit quantised safetensors — runs with mlx-lm, not compatible with LM Studio / llama.cpp / GGUF loaders
Author: Charles Tao — stickmancharles.me | echoforgellc.tech

Requirements

Apple Silicon Mac (M1 / M2 / M3 / M4)
Python ≥ 3.9
mlx-lm package

pip install mlx-lm

Quick start

Interactive chat

mlx_lm.chat --model StickmanCharles/Cardano-Aiken-by-Charles-Tao-v0

Single prompt (CLI)

bash

mlx_lm.generate \
  --model StickmanCharles/Cardano-Aiken-by-Charles-Tao-v0 \
  --prompt "Write an Aiken validator that checks a redeemer equals 42."

Python API

python

from mlx_lm import load, generate

model, tokenizer = load("StickmanCharles/Cardano-Aiken-by-Charles-Tao-v0")

prompt = "Write an Aiken validator that checks a redeemer equals 42."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=text, max_tokens=512, verbose=True)

Repository contents

File / Folder	Description
`model.safetensors`	MLX 4-bit quantised weights
`model.safetensors.index.json`	Weight shard index
`config.json`	Model architecture config (Qwen2, 4-bit)
`tokenizer.json` / `tokenizer_config.json`	Tokeniser files (includes chat template)
`vocab.json` / `merges.txt`	BPE vocabulary
`added_tokens.json` / `special_tokens_map.json`	Special tokens
`aiken_adapters/`	Raw LoRA adapter checkpoints (already fused into main weights)

Notes

This model uses the ChatML prompt format (<\|im_start\|> / <\|im_end\|>).
The LoRA adapters in aiken_adapters/ are already fused into model.safetensors; they are kept for reference only.
For GGUF / LM Studio support, the model would need to be re-converted from the original bf16 weights.