https://github.com/thewh1teagle/gemma3-g2p
Transform Gemma3 to G2P task with unsloth
https://github.com/thewh1teagle/gemma3-g2p
g2p hebrew linguistics tts
Last synced: 5 months ago
JSON representation
Transform Gemma3 to G2P task with unsloth
- Host: GitHub
- URL: https://github.com/thewh1teagle/gemma3-g2p
- Owner: thewh1teagle
- Created: 2025-09-29T03:41:36.000Z (10 months ago)
- Default Branch: main
- Last Pushed: 2025-12-26T23:56:22.000Z (7 months ago)
- Last Synced: 2025-12-27T06:40:29.835Z (7 months ago)
- Topics: g2p, hebrew, linguistics, tts
- Language: Python
- Homepage: https://phonikud.github.io
- Size: 665 KB
- Stars: 3
- Watchers: 0
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# Gemma3-G2P
Based on https://unsloth.ai/blog/gemma3
## Train
See `src/prepare_dataset.py` and `src/train.py`
## Infer
See `src/infer.py`
## Export to GGUF
See `src/export_gguf.py`
## Inference with Ollama
See `ollama-infer/`
## Inference with GGUF
See `gguf-infer/`
## Continue fine-tuning
```console
mkdir outputs
uv run hf download thewh1teagle/gemma3-270m-heb-g2p --local-dir ./outputs/checkpoint-10000
uv run src/prepare_data.py --input_file raw_data.csv --output_file data.csv
uv run src/train.py --report_to tensorboard --csv_file data.csv --resume_from_checkpoint --batch_size 16 --max_steps 90000000
```
## DGX Spark Setup
To fine-tune on DGX Spark, run these commands first:
```console
uv sync
uv pip install -U torch torchvision --index-url https://download.pytorch.org/whl/cu130
export UV_SYNC=0
```