Qwen3-1.7B Issa Haiku (Standalone, Merged)

This is a standalone, merged version of Taku-Yumilab/qwen3-1.7b-issa-haiku-lora. The LoRA adapter weights have been merged into the base Qwen/Qwen3-1.7B weights, so this model can be loaded directly without the peft library.

Intended Output Structure

1. Japanese haiku (Kanji and Kana)
2. Blank line
3. English poetic translation (2-4 lines)

Usage Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Taku-Yumilab/qwen3-1.7b-issa-haiku-standalone"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

messages = [
    {"role": "system", "content": "You're a bilingual haiku poet who writes haiku in Japanese and provides an English translation."},
    {"role": "user", "content": "Compose a haiku about a dragonfly"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.7, top_p=0.8, top_k=20)

print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

Training Summary

Fine-tuned via QLoRA (r=32, alpha=64) for 1 epoch on 10,820 bilingual Issa haiku examples, then merged into the base weights for standalone deployment.