This adapter teaches Nemotron-Nano-4B to produce grounded JSON diagnoses from deterministic audio features + rule-engine candidates. The model never hears raw audio — it reasons over 14 measured acoustic features (spectral centroid, RMS, onset rate, harmonic ratio, etc.) extracted by librosa.

Training

Base model: nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
Dataset: DCASE 2025 Task 2 (real industrial machine audio, 7 machine types)
Training pairs: 300 (feature descriptions + rule candidates → JSON responses)
Epochs: 2
LoRA config: r=16, alpha=32, dropout=0.05, targets: q/k/v/o_proj
Loss convergence: 1.85 → 0.72 (epoch 1, step 100)

Usage with PEFT

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16", torch_dtype="auto"
)
model = PeftModel.from_pretrained(base, "mitvho09/sound-broken-nemotron-lora")
tok = AutoTokenizer.from_pretrained("mitvho09/sound-broken-nemotron-lora")

Context

Part of the "Does It Sound Broken?" app — record 10 seconds of an appliance, get a diagnosis. The full pipeline:

librosa extracts 14 deterministic features (CPU)
Rule engine ranks candidate faults (CPU, transparent)
This LoRA (on Nemotron-4B) produces grounded JSON explanation (GPU)
json_guard validates output grounding (CPU)

The rule engine is the floor; this adapter makes the LLM narration more accurate.