gaokerena/amestris-1b-dpo

SFT-hard-27k Gemma3 English→German Benchmark

This repository contains the supervised fine-tuning benchmark checkpoint for the hard-27k English→German translation experiment.

Repository

[github repository](https://github.com/Mehrdadghassabi/Amestris)

Base model

google/gemma-3-1b-it

Training setup

This benchmark follows the SFT-hard-27k setting:

• Input column: prompt
• Target column: prefered_answer
• Audit-only column: rejected_answer
• Fine-tuning method: LoRA / PEFT
• LoRA rank: 32
• LoRA alpha: 32
• LoRA dropout: 0.05
• Max sequence length: 768
• Epochs: 1
• Learning rate: 5e-05
• Seed: 42

Contents

• sft_hard27k_lora_adapter/: PEFT LoRA adapter checkpoint.
• archives/: compressed archive created after Cell 9A, if uploaded.
• metadata/: upload metadata and reproducibility information.

Notes

The uploaded adapter is intended to be loaded with the gated base model google/gemma-3-1b-it. Users must have accepted the Gemma license on Hugging Face to load the base model.

Bibtex

if you found our model useful feel free to give us a cite!

@misc{amestris-1b-dpo,
  title={Backtranslation Augmented Direct Preference Optimization for Neural Machine Translation},
  author={Ghassabi, Mehrdad and Rajabi, Spehr and Baradaran Kashani, Hamidreza and Hakim, Sadra and Keivandarian, Mahshid and Jahani Bahnamiri, Amirhossein},
  year={2026}
  eprint={2604.25702},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}