Instructions to use STELLiQ/aria-aar-3b-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use STELLiQ/aria-aar-3b-lora with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("unsloth/llama-3.2-3b-instruct-unsloth-bnb-4bit")
model = PeftModel.from_pretrained(base_model, "STELLiQ/aria-aar-3b-lora")

llama-cpp-python

How to use STELLiQ/aria-aar-3b-lora with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="STELLiQ/aria-aar-3b-lora",
	filename="aria-aar-3b-q4_k_m.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use STELLiQ/aria-aar-3b-lora with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf STELLiQ/aria-aar-3b-lora:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf STELLiQ/aria-aar-3b-lora:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf STELLiQ/aria-aar-3b-lora:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf STELLiQ/aria-aar-3b-lora:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf STELLiQ/aria-aar-3b-lora:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf STELLiQ/aria-aar-3b-lora:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf STELLiQ/aria-aar-3b-lora:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf STELLiQ/aria-aar-3b-lora:Q4_K_M

Use Docker

docker model run hf.co/STELLiQ/aria-aar-3b-lora:Q4_K_M

LM Studio
Jan

vLLM

How to use STELLiQ/aria-aar-3b-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "STELLiQ/aria-aar-3b-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "STELLiQ/aria-aar-3b-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/STELLiQ/aria-aar-3b-lora:Q4_K_M

Ollama
How to use STELLiQ/aria-aar-3b-lora with Ollama:
```
ollama run hf.co/STELLiQ/aria-aar-3b-lora:Q4_K_M
```

Unsloth Studio

How to use STELLiQ/aria-aar-3b-lora with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for STELLiQ/aria-aar-3b-lora to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for STELLiQ/aria-aar-3b-lora to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for STELLiQ/aria-aar-3b-lora to start chatting

How to use STELLiQ/aria-aar-3b-lora with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf STELLiQ/aria-aar-3b-lora:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "STELLiQ/aria-aar-3b-lora:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use STELLiQ/aria-aar-3b-lora with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf STELLiQ/aria-aar-3b-lora:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default STELLiQ/aria-aar-3b-lora:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use STELLiQ/aria-aar-3b-lora with Docker Model Runner:
```
docker model run hf.co/STELLiQ/aria-aar-3b-lora:Q4_K_M
```

Lemonade

How to use STELLiQ/aria-aar-3b-lora with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull STELLiQ/aria-aar-3b-lora:Q4_K_M

Run and chat with the model

lemonade run user.aria-aar-3b-lora-Q4_K_M

List all available models

lemonade list

ARIA AAR 3B LoRA — On-Device Meeting Summarization

Fine-tuned Llama 3.2 3B Instruct LoRA adapter for structured meeting summarization, producing TC 7-0.1 After Action Review (AAR) JSON output.

Built for ARIA — an on-device AI meeting assistant running on Samsung Galaxy S24 Ultra (Snapdragon 8 Gen 3).

Model Details

Parameter	Value
Base Model	Llama 3.2 3B Instruct
Method	QLoRA (4-bit NF4)
LoRA Rank	32
LoRA Alpha	32
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Examples	800
Epochs	5
Learning Rate	2e-4 (linear decay)
Max Sequence Length	6144
Final Loss	0.724
Trainable Parameters	~44M / 3.2B (1.4%)

Task Types

The model supports three distinct task types via different system prompts:

1. Single-Pass Summarization

Direct transcript-to-AAR JSON for meetings under ~3,400 words. Produces structured JSON with 6 fields.

2. Chunk Extraction

Extracts structured bullet points (Decisions, Action Items, Key Points, Issues, Notable Quotes) from transcript segments. Used in the hybrid pipeline for long meetings.

3. Refine

Progressive refinement — takes a draft AAR JSON and additional transcript context, produces an improved AAR JSON. Enables processing of arbitrarily long meetings.

Output Format

{
  "title": "Meeting Title in Title Case",
  "what_was_planned": "What was intended to be accomplished...",
  "what_happened": "What actually occurred during the meeting...",
  "why_it_happened": "Analysis of why outcomes differed from plans...",
  "how_to_improve": "Specific actionable recommendations...",
  "ai_perspective": "AI analysis of meeting dynamics and patterns..."
}

Validation Scores

Tested at device-realistic settings: 1536 max tokens, temperature 0.1.

Task	Avg Score	Pass Rate
Brief (< 500 words)	98.4	5/5
Standard (500-1000 words)	93.1	7/7
Detailed (1000-2000 words)	88.6	4/5
Chunk Extraction	77.0	7/10
Refine	100.0	5/5

GGUF

A pre-quantized Q4_K_M GGUF (~1.9GB) is included for direct use with llama.cpp or on-device inference.

Usage

With Transformers + PEFT

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-3B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, "STELLiQ/aria-aar-3b-lora")
tokenizer = AutoTokenizer.from_pretrained("STELLiQ/aria-aar-3b-lora")

With llama.cpp (GGUF)

llama-cli --model aria-aar-3b-q4_k_m.gguf \
  -p "<|start_header_id|>system<|end_header_id|>\n\nYou are an expert meeting analyst...<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nSummarize this meeting transcript:\n\n{transcript}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"

On-Device Performance (Samsung Galaxy S24 Ultra)

Metric	Value
GGUF Size	~1.9 GB (Q4_K_M)
Peak RAM	~2.5 GB
TTFT	~0.5-0.8s (Adreno 750 GPU)
Decode Speed	~50-70 tok/s
GPU Layers	32 (full offload)

Training Data

800 custom examples across three task types:

640 single-pass (brief/standard/detailed tiers)
60 chunk extraction
100 refine (80 from extended transcripts + 20 pilot)

All training data was synthetically generated using meeting transcripts with diverse topics, speaker counts, and meeting styles.

Training Infrastructure

GPU: NVIDIA GeForce RTX 5080 Laptop GPU (16GB)
Framework: Unsloth + Transformers + TRL
Training Time: ~33 minutes
Precision: BFloat16 with 4-bit QLoRA

License

This adapter inherits the Llama 3.2 Community License.

Developed By

STELLiQ Technologies — ARIA: Automated Review Intelligence Assistant

Downloads last month: 6

GGUF

Model size

3B params

Architecture

llama

Hardware compatibility

4-bit

Model tree for STELLiQ/aria-aar-3b-lora

Base model

meta-llama/Llama-3.2-3B-Instruct

Adapter

(757)

this model