CabinLavatoryPrediction

This repository contains a QLoRA adapter fine-tuned from Qwen/Qwen3.5-9B for aircraft lavatory millimeter-wave radar behavior understanding.

The model consumes structured radar time-window information and intermediate representations, then predicts:

• Structured lavatory behavior state: current behavior, transition flag, elapsed/remaining time, next possible behavior, stage index, total stages, and sequence-so-far.
• QA state: occupancy, estimated time to free, used lavatory areas, and abnormal-state flag.

Files

• adapter_model.safetensors, adapter_config.json: final PEFT LoRA adapter.
• checkpoint-6283/: final trainer checkpoint, including optimizer/scheduler state for training resume.
• code/: preprocessing, training, evaluation, visualization, and report scripts.
• eval/metrics/: base vs fine-tuned evaluation metric JSON files.
• eval/charts/: standalone SVG vector charts with embedded metadata.
• presentation/: self-contained design-review HTML/PDF report.

Raw train/validation JSONL data is not included in this model repository.

Training

• Base model: Qwen/Qwen3.5-9B
• Method: 4-bit QLoRA supervised fine-tuning
• LoRA target modules: q/k/v/o/gate/up/down projection modules
• LoRA rank: 16
• LoRA alpha: 32
• Max sequence length used for the successful full run: 2048
• Train data: mixed structured-prediction and QA samples

Evaluation Summary

Validation size:

• Structured task: 4,030 examples
• QA task: 4,030 examples

Key base vs fine-tuned metrics:

Task	Metric	Base	Fine-tuned
Structured	JSON parse rate	98.0%	100.0%
Structured	Required field complete rate	0.0%	95.1%
Structured	Current behavior accuracy	48.1%	67.0%
Structured	Current behavior macro-F1	11.1%	49.1%
Structured	Next possible behavior accuracy	39.2%	65.0%
Structured	Stage index accuracy	0.0%	65.5%
Structured	Sequence exact match	0.0%	61.1%
QA	Occupied accuracy	99.7%	100.0%
QA	Abnormal F1	45.4%	89.5%
QA	Used areas micro-F1	70.5%	100.0%
QA	Time-to-free MAE	5.13 min	~0.0 min

Loading Example

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

base_model = "Qwen/Qwen3.5-9B"
adapter = "sutama/CabinLavatoryPrediction"

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    quantization_config=quant_config,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

Intended Use

This adapter is intended for research and design review of privacy-preserving aircraft lavatory state prediction from structured millimeter-wave radar representations.

It should not be used as an aircraft safety-critical control system without further validation, calibration, monitoring, and fail-safe integration.

Limitations

• The adapter was trained and evaluated on the available structured dataset only.
• Cross-aircraft, cross-installation-angle, sensor-noise, passenger-diversity, and operational robustness require additional validation.
• QA results may partially reflect deterministic target construction rules in the processed dataset; evaluate on independently collected operational data before deployment.