mlx-ssd 0.1.0

Simple Self-Distillation training pipeline for MLX models

mlx-ssd

mlx-ssd is a practical MLX CLI implementation of simple self-distillation for code generation models on Apple Silicon.

Method

This project follows the method introduced in:

Ruixiang Zhang, Richard He Bai, Huangjie Zheng, Navdeep Jaitly, Ronan Collobert, Yizhe Zhang.
Embarrassingly Simple Self-Distillation Improves Code Generation.
arXiv:2604.01193, 2026.
https://arxiv.org/abs/2604.01193

Implementation by Amirani Labs.

Core flow:

1. Sample responses from a base model with train-time decoding settings.
2. Fine-tune on those self-generated samples.
3. Evaluate/run with eval-time decoding settings.

Dataset defaults:

• --problems microsoft/rStar-Coder
• --dataset-config seed_sft
• --dataset-split train
• Records must contain a non-empty question field.

This repository is an independent implementation and is not the original paper repository.

Presets

Presets encode paper-aligned hyperparameters (Table 3 mapping) for supported model families.

mlx-ssd sample --model mlx-community/Qwen3-4B-Instruct-4bit --preset qwen3-4b-instruct --output ./ssd_data
mlx-ssd train --model mlx-community/Qwen3-4B-Instruct-4bit --preset qwen3-4b-instruct --data ./ssd_data --output ./ssd_model
mlx-ssd run --model ./ssd_model/fused --preset qwen3-4b-instruct --prompt "Write a function that..."

Usage

Install:

pip install -e .

Three-stage flow:

# 1) Sample
mlx-ssd sample \
  --model mlx-community/Qwen3-4B-Instruct-4bit \
  --problems microsoft/rStar-Coder \
  --dataset-config seed_sft \
  --dataset-split train \
  --output ./ssd_data \
  --batch-size 16 \
  --temperature 1.6 \
  --top-k 20 \
  --top-p 0.8 \
  --limit 10

# 2) Train
mlx-ssd train \
  --model mlx-community/Qwen3-4B-Instruct-4bit \
  --data ./ssd_data \
  --output ./ssd_model \
  --iters 2500

# 3) Run
mlx-ssd run \
  --model ./ssd_model/fused \
  --temperature 1.1 \
  --top-k 20 \
  --top-p 0.8 \
  --prompt "Write a function that..."

One-command flow:

mlx-ssd distill \
  --model mlx-community/Qwen3-4B-Instruct-4bit \
  --preset qwen3-4b-instruct \
  --output ./my-better-qwen

Local smoke test (quick validation):

mlx-ssd sample \
  --model mlx-community/SmolLM2-135M-Instruct \
  --problems microsoft/rStar-Coder \
  --dataset-config seed_sft \
  --dataset-split train \
  --output ./.smoke/data \
  --batch-size 4 \
  --temperature 0.8 \
  --top-k 20 \
  --top-p 0.8 \
  --max-tokens 64 \
  --limit 5

Apple Silicon

mlx-ssd itself is the Apple Silicon implementation: it is built on mlx-lm and targets local MLX workflows.

License

MIT