Downloads · 30 days
0
hushpond/llama-3-seoul-culture-lora-rag
llama-3-seoul-culture-lora-rag is a machine learning model from hushpond. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as mit.
<h1🎭 Seoul Culture Event Recommendation RAG System (LoRA Adapter)</h1
Downloads · 30 days
0
Access
Public
Updated Dec 13, 2025
Repo size
356 MB
Likes
0
Public
Click a slice to open those files.
.safetensors336 MB · 94%
From the Hugging Face model README
base_model_id = "beomi/Llama-3-Open-Ko-8B" bnb_config = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16 )
base_model = AutoModelForCausalLM.from_pretrained( base_model_id, quantization_config=bnb_config, device_map="auto" ) tokenizer = AutoTokenizer.from_pretrained(base_model_id)
adapter_model_id = "YOUR_HUGGINGFACE_ID/YOUR_MODEL_NAME" model = PeftModel.from_pretrained(base_model, adapter_model_id)
def generate_response(context, persona, query): input_text = f"Context: {context} Persona: [{persona}] 질문: [{query}]" prompt = f"""### Instruction: {persona} 톤으로 답변해 주세요.
{input_text}
""" inputs = tokenizer(prompt, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=300, temperature=0.7, top_p=0.9, eos_token_id=tokenizer.eos_token_id ) return tokenizer.decode(outputs, skip_special_tokens=True).split("### Output:")[-1].strip()
context_data = "[서울 핸드 메이드 페어 2025, 사전 예매: 8,000원, 서울 삼성동 코엑스 1층 B홀]" print(generate_response(context_data, "친절한 문화 가이드", "강남구에서 열리는 핸드메이드 행사 알려줘.")) </code></pre>
<hr> <h2>📊 Training Details</h2> <h3>Dataset</h3> <ul> <li><strong>Source:</strong> 서울열린데이터광장 (문화 행사 정보)</li> <li><strong>Size:</strong> 4,275 rows (Original CSV), Augmented for training</li> <li><strong>Format:</strong> <code>Instruction</code> (페르소나 지시), <code>Input</code> (Context + Question), <code>Output</code> (Answer)</li> <li><strong>Preprocessing:</strong> UTF-8 / CP949 인코딩 자동 처리, 결측치 제거, 512 토큰 길이 제한</li> </ul> <h3>Hyperparameters</h3> <table border="1" cellpadding="5"> <thead> <tr> <th>Parameter</th> <th>Value</th> <th>Reason</th> </tr> </thead> <tbody> <tr> <td><strong>LoRA Rank (r)</strong></td> <td><code>32</code></td> <td>초기 Rank 8/16 실험 결과, Rank 32에서 Loss 1.51로 가장 우수한 성능 기록</td> </tr> <tr> <td><strong>LoRA Alpha</strong></td> <td><code>32</code></td> <td>Scaling factor</td> </tr> <tr> <td><strong>Target Modules</strong></td> <td><code>q_proj</code>, <code>k_proj</code>, <code>v_proj</code>, <code>o_proj</code>, <code>gate_proj</code>, <code>up_proj</code>, <code>down_proj</code></td> <td>모든 Linear Layer 학습으로 추론 능력 극대화</td> </tr> <tr> <td><strong>Batch Size</strong></td> <td><code>1</code></td> <td>Colab T4 GPU 메모리 제약 극복</td> </tr> <tr> <td><strong>Gradient Accumulation</strong></td> <td><code>8</code></td> <td>실제 Batch Size 8 효과 (안정적 수렴)</td> </tr> <tr> <td><strong>Learning Rate</strong></td> <td><code>2e-4</code></td> <td>-</td> </tr> <tr> <td><strong>Epochs</strong></td> <td><code>3</code></td> <td>-</td> </tr> <tr> <td><strong>Optimizer</strong></td> <td><code>paged_adamw_8bit</code></td> <td>메모리 효율 최적화</td> </tr> </tbody> </table> <h3>Performance Metrics</h3> <ul> <li><strong>Validation Loss:</strong> <code>1.5152</code> (Rank 32 기준) <ul> <li>Rank 8 Loss: 2.1565</li> <li>Rank 16 Loss: 1.7950</li> <li><strong>Rank 32 Loss: 1.5152 (Best)</strong></li> </ul> </li> <li><strong>ROUGE-L Score:</strong> <code>0.83</code> (높은 답변 정확도 및 일관성 확인)</li> <li><strong>Temperature Analysis:</strong> Temperature 0.5~0.7 구간에서 ROUGE 점수 최적화 확인 (너무 낮으면 반복, 너무 높으면 환각 발생)</li> </ul> <div align="center"> <table border="1" cellpadding="5"> <tr> <th>Rank</th> <th>Validation Loss</th> </tr> <tr> <td>8</td> <td>2.1565</td> </tr> <tr> <td>16</td> <td>1.7950</td> </tr> <tr> <td><b>32</b></td> <td><b>1.5152</b></td> </tr> </table> </div> <hr> <h2>⚠️ Limitations & Future Work</h2> <ol> <li><strong>Hallucination (환각 현상):</strong> <ul> <li>모델이 학습하지 않은 외부 정보(Unknown Data)에 대해 약 <strong>50%의 환각 발생률</strong>을 보였습니다.</li> <li>RAG 시스템에서 검색되지 않은 정보에 대해 "알 수 없음"이라고 답변하도록 추가 학습(Negative Sampling)이 필요합니다.</li> </ul> </li> <li><strong>Hardware Constraints:</strong> <ul> <li>Colab T4 (free tier) 환경의 제약으로 인해 Batch Size를 1로 설정했습니다. 더 큰 VRAM 환경에서 Full Fine-tuning 시 성능 향상이 기대됩니다.</li> </ul> </li> <li><strong>Scope:</strong> <ul> <li>현재 서울시 데이터에 특화되어 있어, 타 지역 행사나 일반 상식 질문에는 답변 품질이 낮을 수 있습니다.</li> </ul> </li> </ol> <hr> <h2>📜 Citation & License</h2> <ul> <li><strong>License:</strong> Llama 3 Community License</li> <li><strong>Base Model:</strong> <a href="https://huggingface.co/beomi/Llama-3-Open-Ko-8B">beomi/Llama-3-Open-Ko-8B</a></li> </ul>