Trained for 2 epochs on NilanE/ParallelFiction-Ja_En-100k using QLoRA. CPO tune is in-progress.
Input should be 500-1000 tokens long. Make sure to set 'do_sample = False' if using HF transformers for inference, or otherwise set temperature to 0 for deterministic outputs.
Prompt format
"""Translate this from Japanese to English:\n### JAPANESE:\n{source_text}\n### ENGLISH:\n"""
- Downloads last month
- 74