GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
gaokerena
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
gaokerena
Model Tree
Base
google/gemma-3-1b-it
Adapter
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
This is not a merged full model. It contains PEFT/LoRA checkpoint files.
The final adapter files are also placed at the repository root so the adapter can be loaded directly with PEFT.
google/gemma-3-1b-itgaokerena/amestris-1b-dpoLoad the base model first, then attach this adapter/checkpoint with PEFT.
Example:
markdown
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase_model_id = "google/gemma-3-1b-it"adapter_repo_id = "gaokerena/amestris-1b-dpo-sft"tokenizer = AutoTokenizer.from_pretrained(adapter_repo_id)base_model = AutoModelForCausalLM.from_pretrained(base_model_id,device_map="auto",dtype="auto",)model = PeftModel.from_pretrained(base_model, adapter_repo_id)model.eval()
adapter_model.safetensors and adapter_config.jsonlora_adapter/: full copy of the final non-merged LoRA adapter foldertrainer_checkpoints/: trainer checkpoint folders if foundtokenizer/: tokenizer files if savedcheckpoint_package_metadata.json: package metadataarchives/: ZIP archive of this checkpoint packageThis checkpoint corresponds to the experiment:
markdown
Gemma3-1B-IT + Amestris1B-DPO adapter -> SFT on hard-27k
The purpose is to compare against the previous Gemma -> SFT run while keeping the 27K prompt-target training data unchanged.
if you found our model useful feel free to give us a cite!
markdown
@misc{amestris-1b-dpo-sft,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}}