Aashishhhhhhhh/verse-v2-nepali-maithili

VERSE V2 — Nepali, Maithili and English Speech Model

VERSE V2 is a multilingual LoRA/PEFT adapter fine-tuned from Gemma 4 E4B for speech transcription and speech-to-text translation across English, Nepali, and Maithili.

This repository contains the final V2 adapter from training step 12761.

Model details

• Developed by: Aashish Thakuri and Team Northlight
• Model type: Multimodal speech-to-text LoRA/PEFT adapter
• Base model: unsloth/gemma-4-e4b-it-unsloth-bnb-4bit
• Languages: English, Nepali, and Maithili
• Fine-tuning method: LoRA / PEFT
• LoRA rank: 8
• LoRA alpha: 16
• LoRA dropout: 0
• Trainable parameters: 18,350,080
• Approximate trainable share: 0.23%
• Final checkpoint: checkpoint-12761
• Funding: Not publicly disclosed
• License: This adapter is distributed subject to the applicable Gemma terms and the licenses of the source dataset and dependencies.

Supported tasks

Transcription

• English audio → English text
• Nepali audio → Nepali text
• Maithili audio → Maithili text

Speech-to-text translation

• English audio → Nepali text
• English audio → Maithili text
• Nepali audio → English text
• Nepali audio → Maithili text
• Maithili audio → English text
• Maithili audio → Nepali text

This gives nine total language directions:

en → en, ne, mai
ne → ne, en, mai
mai → mai, en, ne

Training data

The source audio dataset was:

Firoj112/chatterbox-multilingual-data

Clean training source clips

Language	Clips
English	21,846
Nepali	25,184
Maithili	19,784
Total	66,814

Each source clip produced one same-language transcription task and two speech-translation tasks.

Final V2 multitask data

Task type	Examples
Transcription	66,814
Speech translation	133,628
Total	200,442

Validation data:

Validation item	Count
Source clips	3,712
Multitask examples	11,136

Why IndicTrans2 was used

The original dataset contained audio and same-language transcripts, but it did not contain every required cross-language translation target.

IndicTrans2 was used only to generate translated text targets for training:

• AI4Bharat/indictrans2-en-indic-200M
• AI4Bharat/indictrans2-indic-en-200M
• AI4Bharat/indictrans2-indic-indic-320M

It was used for English ↔ Nepali, English ↔ Maithili, and Nepali ↔ Maithili target generation. IndicTrans2 did not generate the source audio and is not used as the final inference model.

Training setup

• Cloud platform: RunPod
• Project path: /workspace/verse-v2
• Training method: LoRA / PEFT
• Epochs: 1
• Per-device batch size: 16
• Final optimizer step: 12,761
• Final checkpoint path: /workspace/verse-v2/gemma4-e4b-verse-v2-multitask/checkpoint-12761

Quick validation results

The current results are a small validation estimate using five clips per language.

Language	Samples	WER	CER
English	5	6.72%	3.05%
Nepali	5	8.20%	1.57%
Maithili	5	30.36%	6.45%
Combined	15	12.35%	3.44%

Lower WER and CER are better. These values are preliminary and should not be treated as a final benchmark or state-of-the-art claim.

Loading the adapter

from peft import PeftModel
from unsloth import FastModel

BASE_MODEL = "unsloth/gemma-4-e4b-it-unsloth-bnb-4bit"
ADAPTER_REPO = "Aashishhhhhhhh/verse-v2-nepali-maithili"

base_model, processor = FastModel.from_pretrained(
    model_name=BASE_MODEL,
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_REPO,
    is_trainable=False,
)

model.eval()

Intended uses

VERSE V2 is intended for research and development involving live captions, subtitle generation, podcast transcription, multilingual speech accessibility, Nepali and Maithili speech recognition, and speech-to-text translation.

Limitations

• Maithili transcription currently has higher WER than English and Nepali.
• The reported validation sample contains only 15 clips.
• Translation targets partly depend on IndicTrans2-generated labels.
• Noise, overlapping speakers, music, long audio, rare dialects, and code-switching may reduce accuracy.
• The adapter requires the compatible Gemma 4 E4B base model.
• Independent evaluation is recommended before production or high-stakes use.

Ethical considerations

This model may produce incorrect transcriptions or translations. Outputs should be reviewed before use in legal, medical, financial, safety-critical, or other high-stakes settings. Users are responsible for respecting privacy, consent, copyright, dataset terms, and applicable laws when processing audio.

Model sources

• Base model: unsloth/gemma-4-e4b-it-unsloth-bnb-4bit
• Dataset: Firoj112/chatterbox-multilingual-data
• Project code: https://github.com/AashishThakuri/-Gemma_Margadarshan_hackathon_Team_Northlight

Attribution

VERSE V2 was built using Gemma 4 by Google DeepMind, Unsloth, Hugging Face Transformers and PEFT, IndicTrans2 by AI4Bharat, RunPod cloud infrastructure, and Firoj112/chatterbox-multilingual-data.

Citation

A formal paper citation is not available yet. Until one is published, cite the model repository:

Aashish Thakuri and Team Northlight.
VERSE V2: Nepali, Maithili and English Speech Model.
Hugging Face, 2026.
https://huggingface.co/Aashishhhhhhhh/verse-v2-nepali-maithili

Disclaimer

This repository contains a LoRA/PEFT adapter, not the full merged Gemma base model. To use VERSE V2, load the compatible Gemma 4 E4B base model and then attach this adapter.