saif24565/sopu-coder-fast-mlx-adapters

🚀 SopuCoder Fast (v1.1-candidate) — Fine-Tuned Edition

SopuCoder Fast (v1.1-candidate) is a candidate release of SopuCoder Fast featuring actual parameter adaptations fine-tuned on Apple Silicon Metal using MLX-LM.

---

🔬 Training & Architecture Details

• Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct / mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit
• Method: Low-Rank Adaptation (LoRA)
• Trainable Layers: 16 transformer layers
• LoRA Hyperparameters: Rank: 8, Scale: 20.0, Dropout: 0.0
• Optimizer: Adam (Learning Rate: 5e-5, Batch Size: 1)
• Training Steps: 60 iterations (Loss reduced from 1.560 to 0.118 train / 0.502 validation)
• Training Throughput: ~395 tokens/second during training on Apple Silicon M1 (Metal Unified Memory)

---

⚡ Performance Benchmarks (Apple M1 16GB)

Metric	Upstream Base (1.5B)	SopuCoder Fast (v1.0.0 4B)	SopuCoder Fast (v1.1-candidate)
Inference Velocity	~35 tokens/sec	~22.38 tokens/sec	60.55 - 68.71 tokens/sec
Peak Memory Footprint	~1.5 GB	~3.30 GB	1.09 GB
Context Window	8,192 tokens	16,384 tokens	8,192 tokens
Repetition Artifacts	Low	None	None

---

🛠️ Usage Instructions

1. Running with MLX-LM (Python / Apple Silicon)

# Install MLX-LM
pip install mlx-lm

# Run inference with LoRA adapters
python -m mlx_lm generate \
  --model mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit \
  --adapter-path training/checkpoints/sopu-coder-mlx-adapters \
  --prompt "<|im_start|>system\nYou are SopuCoder Fast.<|im_end|>\n<|im_start|>user\nWrite a React useOptimisticMutation hook.<|im_end|>\n<|im_start|>assistant\n"

2. Testing Locally with Ollama

# Build candidate model
ollama create sopu-coder-fast:v1.1-candidate -f sopu-coder-fast/Modelfile.v1.1-candidate

# Run candidate model
ollama run sopu-coder-fast:v1.1-candidate

---

📜 Provenance & Attribution

• Base Weights: Alibaba Cloud Qwen Team (Qwen/Qwen2.5-Coder-1.5B-Instruct), licensed under Apache 2.0.
• Fine-Tuning Dataset: datasets/v1.1/ curated by SopuCoder open-source contributors (Apache 2.0).
• LoRA Adapter Weights: SopuCoder Project (Apache 2.0).