chim-style-debate-gemma4-e2b-lora-v1

chim-style-debate-gemma4-e2b-lora-v1

This repository contains a LoRA adapter fine-tuned from the base model google/gemma-4-E2B for Korean debate-style responses.

Very important: use the persona prompt

This is an adapter-only repository. The adapter does not behave like a fully packaged chat application by itself. For the intended style, load all three parts together:

1. Base model: google/gemma-4-E2B
2. LoRA adapter: this repository
3. The included persona_system_prompt.txt and chat_template.jinja

If you load the adapter without the persona system prompt, the output can look generic and may not strongly show the intended Korean debate persona.

Important status

This is a v1 / legacy comparison adapter. It was trained before the later v3_relative prompt correction, so it should be treated as a baseline or backup artifact rather than the final recommended adapter.

Known v1 limitation:

• Some training prompts used fixed stance labels such as assistant-pro/user-con for source groups that were later found to be mixed by actual user utterance direction.
• A corrected v3_relative adapter is being trained separately to make rebuttal direction depend on the actual previous user utterance.

Base model requirement

Use this adapter only with:

google/gemma-4-E2B

This project intentionally uses the base Gemma 4 E2B model, not an instruction-tuned variant.

Quick start

from pathlib import Path
import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base_model_id = "google/gemma-4-E2B"
adapter_id = "hyun7120/chim-style-debate-gemma4-e2b-lora-v1"

processor = AutoProcessor.from_pretrained(base_model_id)
model = AutoModelForImageTextToText.from_pretrained(
    base_model_id,
    device_map="auto",
    dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

# You can also read these from the repository files after downloading them.
with open("persona_system_prompt.txt", "r", encoding="utf-8") as f:
    persona_system_prompt = f.read().strip()
with open("chat_template.jinja", "r", encoding="utf-8") as f:
    processor.chat_template = f.read()

messages = [
    {"role": "system", "content": persona_system_prompt},
    {"role": "user", "content": "?? ??: ?? ??? ???? ?? ???? ??? ???
?? ??: ?? ???? ??? ??? ?? ??? ??????.
? ?? ??? ??? ???."},
]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, text_kwargs={"return_tensors": "pt"}).to(model.device)

turn_eot = processor.tokenizer.convert_tokens_to_ids("<turn|>")
eos_ids = [processor.tokenizer.eos_token_id, turn_eot]

with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=180,
        do_sample=True,
        temperature=0.65,
        top_p=0.90,
        top_k=50,
        eos_token_id=eos_ids,
        pad_token_id=processor.tokenizer.eos_token_id,
    )

answer = processor.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=False)
print(answer.split("<turn|>")[0].strip())

For a more complete example, see inference_example.py.

Adapter contents

• adapter_model.safetensors
• adapter_config.json
• persona_system_prompt.txt
• chat_template.jinja
• inference_example.py
• tokenizer / processor files copied for reproducibility

Training summary

• Method: LoRA SFT
• Epochs: 1
• Final global step: 686
• Final train loss: approximately 1.441
• Final eval loss: approximately 1.381
• Final eval mean token accuracy: approximately 0.639

Intended use

• Korean debate response generation
• Baseline comparison against later corrected adapters
• Coursework demonstration and ablation

Limitations

• This adapter is not a general-purpose assistant.
• This adapter is style-conditioned and should be used with the included persona system prompt.
• It may inherit base-model limitations and dataset artifacts.
• It should not be treated as factually reliable without external verification.
• It may produce incomplete or stylistically over-repetitive responses under high-temperature decoding.