---
license: apache-2.0
language:
- en
tags:
- llama
- gguf
- meeting-summarization
- after-action-review
- on-device
pipeline_tag: text-generation
---

# ARIA AAR 1B — Meeting Summarization Model

Fine-tuned **Llama 3.2 1B Instruct** specialized for meeting transcript summarization.
Produces structured JSON output with 5 fields: title, what_was_planned, what_happened, why_it_happened, how_to_improve.

## Model Details

| Property | Value |
|----------|-------|
| Base Model | Llama 3.2 1B Instruct |
| Fine-tuning | QLoRA (r=16, all linear layers) |
| Training Data | 624 examples (real-world + hand-crafted) |
| Quantization | Q4_K_M (4-bit K-quant medium) |
| File Size | 771 MB |
| Context Length | 4096 tokens |
| License | Apache 2.0 |

## Intended Use

On-device meeting summarization for the ARIA Android app.
Runs on mobile GPUs (Adreno, Mali) via llama.cpp with OpenCL acceleration.

## Output Format



## Validation Results

| Test | Score |
|------|-------|
| Average | 98/100 |
| Passing (>=70) | 10/10 |

Built by [STELLiQ](https://github.com/STELLiQ) for the ARIA project.