amestris-1b-dpo-sft is the sequential DPO→SFT checkpoint from the Amestris English-to-German machine-translation study. Training begins from google/gemma-3-1b-it, applies the Amestris DPO adaptation, and then performs supervised fine-tuning on the preferred translations from the same hard-27k research setting.

The released artifact is a non-merged PEFT/LoRA adapter. Although the training procedure is sequential, the final saved adapter can be attached directly to the Gemma base model; users do not need to load the separate DPO adapter first.

Model details

Field	Value
Task	English → German machine translation
Architecture	Gemma 3 1B instruction-tuned causal language model + LoRA
Base model	`google/gemma-3-1b-it`
Training sequence	DPO followed by SFT
Starting DPO release	`gaokerena/amestris-1b-dpo`
Final checkpoint type	PEFT LoRA adapter, non-merged
LoRA rank / alpha / dropout	32 / 32 / 0.05
Maximum sequence length	768 tokens
SFT epochs	1
SFT learning rate	5 × 10⁻⁵
Random seed	42
Languages	English input, German output

The complete methodology is described in the Amestris repository and the associated paper.

Quick start: run the model

1. Install dependencies

pip install -U "transformers>=4.50.0" "peft>=0.18.1" "accelerate>=1.0" torch

Accept Google’s Gemma terms on the base-model page, then authenticate:

hf auth login

2. Load the final adapter and translate

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL_ID = "google/gemma-3-1b-it"
ADAPTER_ID = "gaokerena/amestris-1b-dpo-sft"

# Use the base-model tokenizer, matching the project evaluation workflow.
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    device_map="auto",
    dtype="auto",
)
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

source_text = "The committee will publish its final report next Tuesday."
translation_prompt = f"""You are an expert professional translator from English to German.
Preserve all meaning, nuance, factual details, names, dates, and numbers.
Use natural, fluent, professional standard German.
Do not omit, summarize, add, or explain anything.
Return only the German translation.

Text to translate:
{source_text}"""

messages = [{"role": "user", "content": translation_prompt}]
formatted_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(formatted_prompt, return_tensors="pt")
device = next(model.parameters()).device
inputs = {name: tensor.to(device) for name, tensor in inputs.items()}

with torch.inference_mode():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        num_beams=4,
        early_stopping=True,
        repetition_penalty=1.05,
        no_repeat_ngram_size=3,
    )

generated_ids = output_ids[0, inputs["input_ids"].shape[1]:]
translation = tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
print(translation)

The repository also retains copies of adapter and trainer artifacts for reproducibility, but standard inference should use the root adapter_config.json and adapter_model.safetensors through the repository ID shown above.

Training methodology

The first stage uses preference pairs containing a prompt, a preferred translation, and a rejected lower-quality translation. DPO uses both preferred and rejected responses to improve relative preference alignment. The subsequent SFT stage uses the same standardized prompts but optimizes only against prefered_answer; rejected responses are not included in the supervised loss.

The saved SFT-stage metadata reports one epoch, 1,570 steps, a training loss of 1.7405, 26,091,520 trainable parameters out of 1,025,977,472 total parameters (2.5431%), and an exact prompt → prefered_answer training mapping. LoRA uses rank 32, alpha 32, and dropout 0.05 across the principal linear projections.

Evaluation

The following results are reported by the project on the WMT14 English-to-German test split.

Metric	gemma3-1b-it baseline	amestris-1b-dpo-sft	better direction
BLEU	0.1573	0.1718	↑
COMET22	0.7698	0.8119	↑
COMET-KIWI22	0.7031	0.7790	↑
METEOR	0.3861	0.4325	↑
TER	0.7765	0.7847	↓
chrF++	41.93	45.24	↑

These are project-reported values and were not independently recomputed while preparing this model card. The sequential checkpoint records the strongest reported BLEU, COMET-DA, COMET-KIWI, METEOR, and chrF++ results among the project’s baseline, DPO, SFT, and DPO→SFT configurations, while its reported TER is higher than the baseline.

Intended use

The model is intended for:

English-to-German translation research;
evaluation of sequential preference and supervised post-training;
comparison with the Amestris DPO-only and SFT-only adapters;
PEFT/LoRA reproducibility and ablation studies.

It is not certified for unsupervised production use or for legal, medical, financial, or safety-critical translation.

Limitations and risks

The checkpoint is specialized for English-to-German translation and has not been validated for other directions.
The compact base model may fail on long, domain-specific, ambiguous, or highly structured inputs.
Project analysis reports occasional errors involving names, dates, numbers, entity preservation, hallucination, and sentence alignment.
Higher automatic scores do not guarantee correctness for every example; TER, in particular, is worse than the reported baseline for this configuration.
Biases and safety limitations of the base model and training data may remain.
Use is subject to the Gemma license and gated-model terms.

Repository contents

Root adapter_model.safetensors and adapter_config.json: final inference adapter.
Root tokenizer files: tokenizer snapshot included with the checkpoint package.
lora_adapter/: copy of the final adapter directory and training metrics.
trainer_checkpoints/: trainer checkpoint retained for reproducibility.
tokenizer/: additional tokenizer artifact copy.
checkpoint_package_metadata.json: provenance and packaging metadata.
archives/: archived checkpoint package.

Citation

bibtex

@misc{ghassabi2026backtranslation,
  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}
}

Acknowledgments

This checkpoint was developed as part of the Amestris research project by Mehrdad Ghassabi, Sepehr Rajabi, Hamidreza Baradaran Kashani, Sadra Hakim, Mahshid Keivandarian, and Amirhossein Jahani Bahnamiri.