Fine-Tuning for Your QTH

"My 40m noise floor isn't yours. The stock model is good. A model that's seen my powerline buzz is better."

Target outcome: +2 dB extra effective SNR gain vs the stock model on your noise environment, without regressing on anything else.

v0.5 target

The rfwhisper train fine-tune workflow ships in v0.5. If you're reading this pre-v0.5, the scripts live on a feature branch — see the Fine-tuning milestone.

Step 1 — Record isolated noise (5–10 min)

Point your rig at a clear part of the band where your normal noise is present but no signals are (between 40m calling frequencies, for example), and record to the dummy-load or your rig's audio output:

rfwhisper train record-noise \
  --input "USB Audio CODEC" \
  --rate 48000 \
  --duration 300 \
  --label powerline-evening \
  --out data/my-noise/

Tip: record different times of day and multiple bands. Inverters change with the sun; PLC noise changes with household activity.

Step 2 — Build a fine-tuning dataset (2 min)

rfwhisper train build-dataset \
  --clean  pretrained-clean/        # ships with rfwhisper
  --noise  data/my-noise/
  --out    data/my-train/
  --snr-range -10 20
  --samples 20000

Step 3 — Fine-tune (20–90 min depending on hardware)

On a GPU

rfwhisper train fine-tune \
  --base   deepfilternet3 \
  --data   data/my-train/ \
  --out    runs/my-qth-v1/ \
  --epochs 10 \
  --lr 1e-4 \
  --method lora       # or 'full' for deeper adaptation

Wall time on an RTX 3060: ~45 min for a LoRA fine-tune on 20 000 mixes.

On CPU (laptop / RPi 5)

rfwhisper train fine-tune --base deepfilternet3 --data data/my-train/ \
  --out runs/my-qth-v1/ --epochs 3 --lr 1e-4 --method lora \
  --device cpu --batch 4

Wall time on an i5-12xxx: ~2 hours. RPi 5: ~12 hours for a tiny model.

Step 4 — Evaluate (5 min)

rfwhisper train evaluate --ckpt runs/my-qth-v1/best.pt --eval data/my-eval/

What must pass:

• ≥ +2 dB additional SNR gain on your noise vs the stock DFN3
• No regressions on a held-out general speech+noise set
• CW transient gate green
• FT8 regression gate green

If any gate fails, iterate:

• Increase the dataset size
• Reduce learning rate
• Try a smaller LoRA rank
• Add more diverse noise classes (record at a different time of day)

Step 5 — Export and deploy

rfwhisper train export --ckpt runs/my-qth-v1/best.pt \
  --out ~/.cache/rfwhisper/models/my-qth-v1/ \
  --variants fp32 int8

Use it:

rfwhisper denoise-live --model my-qth-v1 --profile ssb

Step 6 — (Optional) Share it

If your fine-tune beats the stock model on your QTH and passes all gates, please consider contributing it to the community hub:

1. Upload the .onnx to a public host (or open a PR with it via Git LFS).
2. Fill out the model card template.
3. Open a PR in the rfwhisper-models repo (or the main repo pre-hub-launch).

We never require you to share. But a fine-tune trained on rural-Wisconsin-solar-inverter-on-40m-evening noise is priceless for the next ham in a similar environment.

LoRA vs full fine-tune

Method	When	Wall time	Quality ceiling
LoRA (rank 8–16)	Default. Your noise is a variant of typical QRM.	Fast	+1–3 dB over stock
Full	Your noise is dramatically unusual (custom RF environment, specific interferer).	5–10× slower	Higher, but risks overfitting / regressions

Start with LoRA. Escalate only if you need to.

Troubleshooting

• Regression gates fail mid-training — lower the learning rate; you're overfitting to the new noise at the cost of signal preservation.
• Tiny SNR gain improvement — your stock model was already great at your noise. That's a win.
• Training diverges — check the noise recordings for clipping or DC offset; rfwhisper train record-noise reports both but it's easy to miss.