Instructions to use Junping0645/vtuber with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use Junping0645/vtuber with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("EleutherAI/polyglot-ko-3.8b") model = PeftModel.from_pretrained(base_model, "Junping0645/vtuber") - Transformers
How to use Junping0645/vtuber with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Junping0645/vtuber")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("Junping0645/vtuber", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Junping0645/vtuber with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Junping0645/vtuber" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Junping0645/vtuber", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/Junping0645/vtuber
- SGLang
How to use Junping0645/vtuber with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Junping0645/vtuber" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Junping0645/vtuber", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Junping0645/vtuber" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Junping0645/vtuber", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use Junping0645/vtuber with Docker Model Runner:
docker model run hf.co/Junping0645/vtuber
QLoRA adapter: 사투리 시드 + 멀티턴 보강 (eval_loss 1.366)
Browse files- README.md +76 -1
- adapter_config.json +43 -0
- adapter_model.safetensors +3 -0
- tokenizer.json +0 -0
- tokenizer_config.json +16 -0
README.md
CHANGED
|
@@ -1,3 +1,78 @@
|
|
| 1 |
---
|
| 2 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
base_model: EleutherAI/polyglot-ko-3.8b
|
| 3 |
+
library_name: peft
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
+
language:
|
| 6 |
+
- ko
|
| 7 |
+
tags:
|
| 8 |
+
- base_model:adapter:EleutherAI/polyglot-ko-3.8b
|
| 9 |
+
- lora
|
| 10 |
+
- transformers
|
| 11 |
+
- korean
|
| 12 |
+
- chatbot
|
| 13 |
---
|
| 14 |
+
|
| 15 |
+
# 말동무 — 독거 어르신용 AI 대화 챗봇 (QLoRA 어댑터)
|
| 16 |
+
|
| 17 |
+
`EleutherAI/polyglot-ko-3.8b`를 QLoRA(4bit NF4)로 파인튜닝한 LoRA 어댑터입니다.
|
| 18 |
+
독거 어르신과 짧고 따뜻한 존댓말로 대화하며, 먼저 공감하고 필요시 안부를 챙기는 페르소나로 학습했습니다.
|
| 19 |
+
|
| 20 |
+
## 학습 데이터
|
| 21 |
+
- 단발 대화 11,266개: 16개 카테고리(건강/외로움/손주/명절 등) × 8개 감정, GPT-4o-mini로 생성 + 실제 어르신 인터뷰 전사에서 추출한 사투리/구어체 발화 1,266개(사투리 입력 이해 보강용)
|
| 22 |
+
- 멀티턴 대화 1,000편(대화당 평균 7.1턴) → 턴마다 슬라이딩 윈도우로 확장해 3,546개 학습 샘플로 변환
|
| 23 |
+
- 최종 train 14,072 / val 740
|
| 24 |
+
|
| 25 |
+
## 학습 설정
|
| 26 |
+
- LoRA: r=8, alpha=16, dropout=0.1, target_modules=["query_key_value", "dense"] (GPT-NeoX 어텐션 위주로 좁게 타깃 — 과적합 억제)
|
| 27 |
+
- 4bit NF4 양자화(bitsandbytes), MAX_LENGTH=768
|
| 28 |
+
- batch_size=2, gradient_accumulation_steps=8 (유효 배치 16), 5 epoch 캡 + EarlyStoppingCallback(patience=3)
|
| 29 |
+
- 결과: epoch 3.296에서 early stopping, **eval_loss 1.366**, RTX 3060 Ti(8GB)에서 141.9분 소요, peak VRAM 3.4GB
|
| 30 |
+
|
| 31 |
+
## 사용법
|
| 32 |
+
|
| 33 |
+
베이스 모델(7.6GB)을 4bit로 로드한 뒤 이 어댑터를 얹습니다.
|
| 34 |
+
|
| 35 |
+
```python
|
| 36 |
+
import torch
|
| 37 |
+
from peft import PeftModel
|
| 38 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
| 39 |
+
|
| 40 |
+
BASE_MODEL = "EleutherAI/polyglot-ko-3.8b"
|
| 41 |
+
ADAPTER = "Junping0645/Vtuber" # 이 리포
|
| 42 |
+
|
| 43 |
+
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
|
| 44 |
+
bnb = BitsAndBytesConfig(
|
| 45 |
+
load_in_4bit=True,
|
| 46 |
+
bnb_4bit_quant_type="nf4",
|
| 47 |
+
bnb_4bit_compute_dtype=torch.float16,
|
| 48 |
+
bnb_4bit_use_double_quant=True,
|
| 49 |
+
)
|
| 50 |
+
base = AutoModelForCausalLM.from_pretrained(BASE_MODEL, quantization_config=bnb, device_map={"": 0})
|
| 51 |
+
model = PeftModel.from_pretrained(base, ADAPTER)
|
| 52 |
+
model.eval()
|
| 53 |
+
|
| 54 |
+
prompt = "### 어르신: 요즘 많이 외롭습니다.\n### 말동무:"
|
| 55 |
+
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
| 56 |
+
out = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.8,
|
| 57 |
+
top_p=0.9, repetition_penalty=1.15,
|
| 58 |
+
pad_token_id=tokenizer.eos_token_id, eos_token_id=tokenizer.eos_token_id)
|
| 59 |
+
print(tokenizer.decode(out[0], skip_special_tokens=True))
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
### 멀티턴(대화 이력 유지)
|
| 63 |
+
|
| 64 |
+
이전 턴들을 같은 포맷으로 이어붙이면 문맥을 반영한 답변이 나옵니다(학습 시 사용한 포맷과 동일해야 함):
|
| 65 |
+
|
| 66 |
+
```
|
| 67 |
+
### 어르신: 손주가 다음 주에 놀러 온대요.
|
| 68 |
+
### 말동무: 정말요? 기대되시겠어요! 즐거운 시간 보내시길 바래요.
|
| 69 |
+
### 어르신: 몇 살인지 안 물어보셨네요, 이번에 초등학교 들어가요.
|
| 70 |
+
### 말동무: 그렇군요! 요즘 아이들은 빠르죠.
|
| 71 |
+
### 어르신: 걔가 오면 뭘 해주면 좋아할까요?
|
| 72 |
+
### 말동무:
|
| 73 |
+
```
|
| 74 |
+
|
| 75 |
+
## 참고
|
| 76 |
+
- VRAM: 학습 3.4GB / 추론 약 3.3GB (RTX 3060 Ti 8GB 기준)
|
| 77 |
+
- Framework: PEFT 0.19.1, transformers 5.13.0, bitsandbytes 0.49.2, torch 2.11.0+cu128
|
| 78 |
+
- 개인 프로젝트용 비공개 모델입니다.
|
adapter_config.json
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alora_invocation_tokens": null,
|
| 3 |
+
"alpha_pattern": {},
|
| 4 |
+
"arrow_config": null,
|
| 5 |
+
"auto_mapping": null,
|
| 6 |
+
"base_model_name_or_path": "EleutherAI/polyglot-ko-3.8b",
|
| 7 |
+
"bias": "none",
|
| 8 |
+
"corda_config": null,
|
| 9 |
+
"ensure_weight_tying": false,
|
| 10 |
+
"eva_config": null,
|
| 11 |
+
"exclude_modules": null,
|
| 12 |
+
"fan_in_fan_out": false,
|
| 13 |
+
"inference_mode": true,
|
| 14 |
+
"init_lora_weights": true,
|
| 15 |
+
"layer_replication": null,
|
| 16 |
+
"layers_pattern": null,
|
| 17 |
+
"layers_to_transform": null,
|
| 18 |
+
"loftq_config": {},
|
| 19 |
+
"lora_alpha": 16,
|
| 20 |
+
"lora_bias": false,
|
| 21 |
+
"lora_dropout": 0.1,
|
| 22 |
+
"lora_ga_config": null,
|
| 23 |
+
"megatron_config": null,
|
| 24 |
+
"megatron_core": "megatron.core",
|
| 25 |
+
"modules_to_save": null,
|
| 26 |
+
"peft_type": "LORA",
|
| 27 |
+
"peft_version": "0.19.1",
|
| 28 |
+
"qalora_group_size": 16,
|
| 29 |
+
"r": 8,
|
| 30 |
+
"rank_pattern": {},
|
| 31 |
+
"revision": null,
|
| 32 |
+
"target_modules": [
|
| 33 |
+
"query_key_value",
|
| 34 |
+
"dense"
|
| 35 |
+
],
|
| 36 |
+
"target_parameters": null,
|
| 37 |
+
"task_type": "CAUSAL_LM",
|
| 38 |
+
"trainable_token_indices": null,
|
| 39 |
+
"use_bdlora": null,
|
| 40 |
+
"use_dora": false,
|
| 41 |
+
"use_qalora": false,
|
| 42 |
+
"use_rslora": false
|
| 43 |
+
}
|
adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:663e23f4d7ce13500be6a1c04b00afab542b482407308c9ac3df78cda556f524
|
| 3 |
+
size 18892264
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backend": "tokenizers",
|
| 3 |
+
"eos_token": "<|endoftext|>",
|
| 4 |
+
"extra_special_tokens": [
|
| 5 |
+
"<|endoftext|>",
|
| 6 |
+
"<|sep|>",
|
| 7 |
+
"<|acc|>",
|
| 8 |
+
"<|tel|>",
|
| 9 |
+
"<|rrn|>"
|
| 10 |
+
],
|
| 11 |
+
"is_local": false,
|
| 12 |
+
"local_files_only": false,
|
| 13 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 14 |
+
"pad_token": "<|endoftext|>",
|
| 15 |
+
"tokenizer_class": "TokenizersBackend"
|
| 16 |
+
}
|