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

json

{
  "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

python

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)

bash

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