GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
aselbaekki
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
aselbaekki
Model Tree
Base
Qwen/Qwen2.5-3B
Fine-tuned
this model
Input Modalities
Text
Output Modalities
Text
Supported Functionality
Dedicated Endpoints
Container
Qwen2.5-3B Korean Petition 100K is a supervised fine-tuned language model based on Qwen2.5-3B. The model was trained on approximately 100,000 Korean public petition samples to improve its ability to understand and generate Korean formal writing, public opinions, and petition-style text.
The model is intended for research purposes and experimentation with Korean language generation and instruction-following tasks.
The model can be used for
Possible downstream applications include
This model should not be used for
The model inherits limitations from the original Qwen2.5-3B model and the petition dataset used during fine-tuning.
Potential limitations include:
Users should verify important outputs before use.
python
from transformers import AutoTokenizer, AutoModelForCausalLMmodel_name = "aselbaekki/Qwen2.5-3B-korean-petition-100k"tokenizer = AutoTokenizer.from_pretrained(model_name)model = AutoModelForCausalLM.from_pretrained(model_name,torch_dtype="auto",device_map="auto")prompt = "대한민국 교육 정책에 대해 설명해주세요."messages = [{"role": "user", "content": prompt}]text = tokenizer.apply_chat_template(messages,tokenize=False,add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs,max_new_tokens=512)print(tokenizer.decode(outputs[0], skip_special_tokens=True))
The model was fine-tuned on approximately 100,000 Korean petition samples.
The dataset consists primarily of Korean public petitions containing formal writing and citizen opinions.
| Parameter | Value |
|---|---|
| Base Model | Qwen2.5-3B |
| Epochs | 1 |
| Learning Rate | 5e-5 |
| Batch Size | 4 |
| Gradient Accumulation | 4 |
| Scheduler | Cosine |
| Warmup Ratio | 0.03 |
| Precision | BF16 |
| Optimizer | AdamW |
The model was qualitatively evaluated on Korean instruction-following and text generation tasks.
Future work will include standardized benchmarks for
If you use this model in your research, please cite both the original Qwen2.5 paper and this repository.
bibtex
@misc{qwen25-korean-petition-100k,author = {Asel Nurlanbek kyzy},title = {Qwen2.5-3B Korean Petition 100K},year = {2026},publisher = {Hugging Face},howpublished = {\url{https://huggingface.co/aselbaekki/Qwen2.5-3B-korean-petition-100k}}}
For questions or feedback, please open an issue on the Hugging Face repository or contact the repository owner.
Hugging Face: https://huggingface.co/aselbaekki
| Framework | Hugging Face Transformers |