VERSE

Three languages.
One trained adapter.

GemmaLoRA / PEFTPyTorchMultilingual speech

Team Northlight’s Gemma fine-tuning project for English, Nepali, and Maithili. A closer look at the data, training setup, and measured error rates behind the model.

The training

V2 continues from a V1 LoRA checkpoint on a 4-bit Gemma 4 E4B base. Rank 8, alpha 16, and 18.35 million trainable parameters keep the update compact. The published run used one epoch, batch size 16, and 12,761 optimizer steps.

The data

66,814 cleaned audio clips become 200,442 training examples: one transcription and two translations per clip. English, Nepali, and Maithili share the same adapter. IndicTrans2 supplies missing translation targets during data preparation.

The performance

The repository reports a matched comparison on 200 held-out clips: WER falls from 41.96% to 18.92%, and CER from 12.48% to 4.08%. A separate V2 check uses only 15 clips; its results are preliminary and are shown separately.

HuggingFace GitHubSource

Matched evaluation: original Gemma WER 41.96%, fine-tuned 18.92%; original CER 12.48%, fine-tuned 4.08%. Lower error is better. — 1 / Matched evaluation200 held-out clips

THE TRAINING SETUP / V2

A small update.
A multilingual objective.

Trainable parameters: 18.35M
Optimizer steps: 12,761
Training epochs: 1

1. PrepareClean audio, retain transcripts, generate missing translation targets.

2. Fine-tuneContinue the V1 adapter with a shared transcription and translation objective.

3. EvaluateDecode held-out speech and report word and character error rates.

Base	Gemma 4 E4B / Unsloth / 4-bit
LoRA	Rank 8 · alpha 16 · dropout 0
Batch size	16 / gradient accumulation 1
Learning rate	0.0002 / warmup ratio 0.03
Compute	Colab A100 and RunPod B200, as reported by the project

The published Hugging Face artifact is a PEFT adapter and requires its compatible base model. Training-loss histories are not published in this repository. Training script ↗

2 / Model trainingGemma + LoRA

Clean training clips: English 21,846, Nepali 25,184, Maithili 19,784. Total 66,814. — 3 / Training dataEnglish / Nepali / Maithili

V2 preliminary validation, five clips per language. WER: English 6.72%, Nepali 8.20%, Maithili 30.36%. CER: 3.05%, 1.57%, 6.45% respectively. — 4 / V2 quick validation5 clips per language

A small update.A multilingual objective.

A small update.
A multilingual objective.