HwanChang0106/qwen3-seq-sort-ko

Qwen3-8B LoRA: Korean Sequence Sorting

This repo contains only the LoRA adapter and tokenizer/template files for a Qwen3-8B base model. Use it by loading the base model first and then applying this adapter with PEFT.

Usage

• Base model: Qwen/Qwen3-8B (replace with your exact base if different)
• Adapter repo (this repo): HwanChang0106/qwen3-seq-sort-ko

Python example:

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

base_id = "Qwen/Qwen3-8B"  # base model
lora_id = "HwanChang0106/qwen3-seq-sort-ko"  # this repo

tok = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, lora_id)

prompt = "다음 수열을 오름차순으로 정렬하세요: 3, 1, 2"
inputs = tok(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64)
print(tok.decode(outputs[0], skip_special_tokens=True))

If you prefer to merge the LoRA into the base weights for pure HF usage:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(base, lora_id)
merged = model.merge_and_unload()  # returns a merged base model

Files

• adapter: adapter_model.safetensors, adapter_config.json
• tokenizer/template: tokenizer.json, tokenizer_config.json, special_tokens_map.json, added_tokens.json, vocab.json, merges.txt, chat_template.jinja

Optimizer or training state files are not included.

Notes

• Ensure trust_remote_code=True when using Qwen models.
• Match the exact base checkpoint you trained on for best results.