muonai/PULSE-2B-LoRa-adapters

PULSE-2B-LoRA

PULSE-2B is a lightweight, high-performance fine-tuned version of Google's Gemma 2 2B Instruct model. Trained locally on Apple Silicon using Low-Rank Adaptation (LoRA) via MLX, this adapter improves instruction-following and system behaviors with minimal memory usage.

• Base Model: mlx-community/gemma-2-2b-it-4bit
• Framework: MLX / Apple Silicon Unified Memory
• Training Iterations: 600
• Final Loss: 2.53

---

🛠️ Requirements & Installation

Make sure you are on a Mac with Apple Silicon (M1/M2/M3/M4) running macOS.

pip install -U mlx-lm huggingface_hub

---

🚀 Quickstart Tutorials

Option 1: Using MLX Command Line Interface (CLI)

1. Text Generation

You can run generation directly without fusing the model weights:

mlx_lm.generate \
  --model mlx-community/gemma-2-2b-it-4bit \
  --adapter-path YOUR_USERNAME/PULSE-2B-LoRA \
  --prompt "Instruction:\nGive me 3 practical tips for time management.\n\nResponse:\n" \
  --max-tokens 200

2. Interactive Terminal Chat

Launch an interactive session directly in your terminal:

mlx_lm.chat \
  --model mlx-community/gemma-2-2b-it-4bit \
  --adapter-path YOUR_USERNAME/PULSE-2B-LoRA

---

Option 2: Using Python API (mlx_lm)

You can easily load the base model and adapter inside Python scripts or Jupyter Notebooks:

from mlx_lm import load, generate

# 1. Load the base quantized model alongside the LoRA adapter
model, tokenizer = load(
    "mlx-community/gemma-2-2b-it-4bit",
    adapter_path="YOUR_USERNAME/PULSE-2B-LoRA"
)

# 2. Format the prompt
prompt = "Instruction:\nExplain quantum computing in simple terms.\n\nResponse:\n"

# 3. Generate response
response = generate(
    model, 
    tokenizer, 
    prompt=prompt, 
    max_tokens=250, 
    verbose=True
)

print(response)

---

Option 3: Using Hugging Face peft & transformers (PyTorch)

If you converted or adapted these weights for standard PyTorch/Transformers workflows, you can load them with peft:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "google/gemma-2-2b-it"
adapter_id = "YOUR_USERNAME/PULSE-2B-LoRA"

# 1. Load Tokenizer & Base Model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# 2. Attach LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_id)

# 3. Generate
inputs = tokenizer("Instruction:\nWrite a 4-line poem about rain.\n\nResponse:\n", return_tensors="pt").to("cuda" if torch.cuda.is_available() else "cpu")
outputs = model.generate(**inputs, max_new_tokens=100)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📜 Notice & License

PULSE-2B is a derivative work based on Google's Gemma 2 weights and is governed by the official Gemma Terms of Use.