
Google Colabã§Text-to-LoRAã詊ã
1. ã¯ããã«
æè¿ãSakana AIããText-to-LoRAãšããè«æãã³ãŒããã¢ãã«ãšå
±ã«çºè¡šãããŠé¢çœãã£ãã®ã§è©ŠããŠã¿ãŸããã
ã³ãŒãã¯å
¬éããã€ã€ãã誰ã§ã詊ããããGoogle Colabã§ã®å®è¡ã«ã¯ïŒäž»ã«Flash Attentionåšãã§ïŒè¥å¹²æéåã£ãããåå¿é²ãšããŠèšäºãæžããŠããããšæããŸãã
2. Text-to-LoRAã®æŠèŠ
å
ã
ãç¹ã«ããŒã«ã«LLMã®é åã§èšç®è³æºãæããŠç¹å®ã¿ã¹ã¯åãã«ãã¥ãŒãã³ã°ããææ³ãšããŠLoRA(Low Rank Adaptation)ã2023幎é ããæ³šç®ãããŠããŸããã
Full FineTuningã«ã¯åã°ãªããªããããå
ã®ã¢ãã«ãå€ããã«AdapterãäœæããŠã¿ã¹ã¯ã«ç¹åã§ããã®ãé
åã§ããã
äžæ¹ã§ãçµå±ã®ãšãããã®ã¿ã¹ã¯ã«ç¹åããããã«ã¯ããŒã¿ãçšæããªããŠã¯ãªããªãç¹ã§ããŸãæ¥åã§ã¯äœ¿ãããŠããªãå°è±¡ã§ãããïŒããããç¹æ®µå¶çŽããªããã°æ¥åã§LLMã䜿ãæã¯ãŸãAzure OpenAI Serviceãªã©ã®APIããå§ããã®ãäžè¬çãªæ°ãããŸããïŒ
Text-to-LoRAã§ã¯ãã¿ã¹ã¯ã®èª¬ææãããLLMãç¹å®ã®ã¿ã¹ã¯ã«ç¹åãããLoRA Adapterãäœæããããšãã§ããŸãã

ææ³ãšããŠã¯ãéåžžã®LoRAã§ã¯AdapterãšãªãÎWã®éã¿ãæŽæ°ããŠãããšããããã¿ã¹ã¯èª¬æ(Task emb)ããAdapter(ÎW)ãçæãããããã¯ãŒã¯(Hypernetwork)ã®éã¿ãæŽæ°ããŸãã
ãããè€æ°ã®ã¿ã¹ã¯ã§è¡ã£ããšããããŒãã·ã§ããé©å¿ãå¯èœã«ãªã£ããšããããšã¿ããã§ããïŒã€ãŸããããçšåºŠæªç¥ã®ã¿ã¹ã¯çšã®LoRA Adapterãçæã§ããããã«ãªã£ããïŒ
3. Google Colabã§è©Šã
åºæ¬çãªåããæ¹ã¯å
¬éãããŠããã³ãŒãã®Readmeã«èšèŒãããŠããŸãã®ã§ãããåç
§ããŸããã
ç§ã¯Google Colabã§å®è¡ããã®ã§ã以äžã®ã³ãŒãã¯ããããåã»ã«ã®å
容ã ã§ããã€ã³ã¹ã¿ã³ã¹ã¯A100ã䜿ããŸããã
3.1 æºå
ãŸãã¯Githubããã³ãŒããååŸããŸãã
# Text-to-LoraååŸ
!git clone https://github.com/SakanaAI/text-to-lora.git
%cd text-to-lora
# ããã±ãŒãžã€ã³ã¹ããŒã«
!pip -q install -e .Google Colabã§ã¯numpyã®ããŒãžã§ã³é¢é£ã§åŸã®ã³ãŒãã§ãšã©ãŒãåºãããnumpyãå ¥ãçŽããŸãããinstallã®åŸã«ã©ã³ã¿ã€ã ãåèµ·åããŸãã
# numpyã®ããŒãžã§ã³ã解決(ãã®åŸãã©ã³ã¿ã€ã åèµ·å)
!pip uninstall -y numpy
!pip install numpyåèµ·ååŸããã£ã¬ã¯ããªã倿Žã
# åèµ·ååŸ
%cd text-to-loraç°å¢å€æ°ãæºåããŠhuggingface-cliã«ãã°ã€ã³ããHuggingFaceããText-to-LoRAã®ã¢ãã«ãååŸããŸãã
# ç°å¢å€æ°ã®æºå
import os
from google.colab import userdata
os.environ["HF_TOKEN"] = userdata.get("HF_TOKEN")!huggingface-cli login --token $HF_TOKEN
!huggingface-cli download SakanaAI/text-to-lora --local-dir . --include "trained_t2l/*"3.2 Flash Attentionã®ã€ã³ã¹ããŒã«
Flash Attentionã®ããŒãžã§ã³ã確èªãinstallããŸããã©ã®Flash Attentionãå
¥ããã°ãããã¯Pytorch, cuda, Pythonã®ããŒãžã§ã³ã«ãã£ãŠæ±ºãŸããŸãã
ãŸããããã確èªããŸãã
!pip show torch
!nvcc -V
!python --version2025/06/29ã®æç¹ã§ãPytorch: v2.4.0, cuda: 12.5, Python: 3.11.13ã§ããã
Flash Attentionã®ææ°ã¯v2.8.0ã ã£ãã®ã§ããã®äžããé©åãªwheelãã¡ã€ã«ãéžæããŠã€ã³ã¹ããŒã«ããŸãã
Flash Attentionã®ãªãªãŒã¹ç¶æ³ã«ã€ããŠã¯ä»¥äžã§ç¢ºèªã§ããŸãã
https://github.com/Dao-AILab/flash-attention/releases
ãã¡ã€ã«åã«Pytorch, cuda, Pythonã®ããŒãžã§ã³ãèšèŒãããŠããŸããabiã¯FALSEã§è¯ãã¿ããã§ãã
# cu125, torch2.4, abiFALSE, cp311
# -> v2.8.0.post2/flash_attn-2.8.0.post2+cu12torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl
# !pip uninstall -y flash-attn
!pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.0.post2/flash_attn-2.8.0.post2+cu12torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl3.3 Text-to-LoRAã§Adapterã®äœæ
generate_lora.pyã䜿ã£ãŠAdapterãäœæããŸããããŒã¹ã¢ãã«ã«ã¯Mistral 7Bã䜿ããŸãããïŒmistral_7b_t2lãæå®ãïŒ
ããã«ãã£ãŠtrained_t2l/mistral_7b_t2l/extras/user_generated/äžã«AdapterãäœæãããŸãã
TASK_DESC="This task requires generating helpful, empathetic customer service responses to complaints or inquiries. Responses should acknowledge the customer's issue, express understanding, provide a clear solution or next steps, and maintain a professional yet friendly tone. The response should be concise (2-3 sentences maximum) and action-oriented. Always prioritize customer satisfaction and problem resolution."
!python scripts/generate_lora.py \
trained_t2l/mistral_7b_t2l \
"$TASK_DESC"ã¿ã¹ã¯èª¬æãšããŠä»åã¯ä»¥äžã®ãããªã¿ã¹ã¯ã«ããŠã¿ãŸããã詊ãã¿ã¹ã¯ã¯Claudeã«äœæããŠããããŸããã
ãã®ã¿ã¹ã¯ã§ã¯ãèŠæ ãåãåããã«å¯ŸããŠã圹ç«ã€ãã€å ±æçãªã«ã¹ã¿ããŒãµãŒãã¹å¯Ÿå¿ãçæããå¿ èŠããããŸãã察å¿ã§ã¯ã顧客ã®èª²é¡ãèªããçè§£ã瀺ãããšãæç¢ºãªè§£æ±ºçãæ¬¡ã®ã¹ããããæäŸããããšããããŠãããã§ãã·ã§ãã«ã§ãããªãã芪ãã¿ãããããŒã³ãä¿ã€ããšãéèŠã§ãã察å¿ã¯ç°¡æœïŒæå€§2ïœ3æïŒã§ãè¡åæåã§ããå¿ èŠããããŸããåžžã«é¡§å®¢æºè¶³åºŠãšåé¡è§£æ±ºãæåªå ããŠãã ããã
3.4 ããŒã¹ã¢ãã«ã§ã®åºåã®ç¢ºèª
ããŒã¹ã¢ãã«ãšãªãmistralai/Mistral-7B-Instruct-v0.2ãååŸããã¿ã¹ã¯ãæããŠã¿ãŸãã
# Mistral-7B ã®ããŒã¹ã©ã€ã³ãçšæ
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
BASE = "mistralai/Mistral-7B-Instruct-v0.2"
tokenizer = AutoTokenizer.from_pretrained(BASE)
base_model = AutoModelForCausalLM.from_pretrained(
BASE,
device_map="auto"
)PROMPT = """Customer complaint: I ordered a laptop 2 weeks ago and it still hasn't arrived. The tracking number doesn't work and nobody responds to my emails. This is completely unacceptable for a premium service!
Response:"""
def generate(model, text):
tokens = tokenizer(text, return_tensors="pt").to("cuda")
out = model.generate(**tokens, max_new_tokens=512)
return tokenizer.decode(out[0], skip_special_tokens=True)
baseline_answer = generate(base_model, PROMPT)
print("â ããŒã¹ã©ã€ã³:\n", baseline_answer)æããã¿ã¹ã¯ã¯ä»¥äžã®éãã§ãã
顧客ã®èŠæ ïŒ2é±éåã«ããŒãããœã³ã³ã泚æããŸãããããŸã å±ããŠããŸããã远跡çªå·ãæ©èœãããã¡ãŒã«ãžã®è¿ä¿¡ããããŸããããã¬ãã¢ã ãµãŒãã¹ãšããŠãããã¯å®å šã«åãå ¥ããããŸããïŒ
åçïŒ
çµæã¯ä»¥äžã®ããã«ãªããŸãããMistralèªäœæ§èœãããããè¯ãã®ã§åçã¯äœæã§ããŠããŸãããä»åè¿ããŠæ¬²ããã察å¿ã¯ç°¡æœïŒæå€§2ïœ3æïŒãã§ãšããã®ã«ã¯æ²¿ã£ãŠããŸãããïŒããã³ããã«å ¥ã£ãŠããªãã®ã§åœç¶ã§ãããïŒ
[ã客æ§]æ§ã
ããŒãããœã³ã³ã®é éã«é¢ãããäžäŸ¿ããããããèª ã«ç³ãèš³ããããŸããããæ³šæã®ã¹ããŒã¿ã¹ã«ã€ããŠæç¢ºãªåçãåŸãããªãããšã¯ãéåžžã«ã¹ãã¬ã¹ãã«ãªç¶æ³ã§ããããšãçè§£ããŠãããŸãã
ãŸããå éšã§æ³šæç¶æ³ã確èªããã§ããã ãæ©ããå ±åããããŸãããã®éãæ³šæçªå·ãšæ³šæã«çŽä»ããã¡ãŒã«ã¢ãã¬ã¹ããæäŸããã ããŸãã§ããããïŒããã«ãããæ å ±ãããå¹ççã«ç¢ºèªã§ããŸãã
ãŸãããã®åé¡ãç©æµããŒã ã«ãšã¹ã«ã¬ãŒã·ã§ã³ããç¶æ³ã確èªããã§ããã ãæ©ã解決ã§ããããåªããŸãã
ãæ³šææã«ãæåŸ ãããŠãããã®ãšã¯ç°ãªãé å»¶ãçºçããŠããããšãçè§£ããŠããããå¿é ããããããŠããããšããè©«ã³ç³ãäžããŸããç§ãã¡ã¯ãèŠæãçå£ã«åãæ¢ãããæºè¶³ããã ããããåªããŠãŸãããŸãã
æ°æéã»ã©ãæéãããã ãã詳现ã確èªãããŠããã ããŸãããã®åŸãéããã«ãé£çµ¡ããããŸãããã®éããäžæç¹ããå¿é äºãããããŸãããããæ°è»œã«ç§ã®ã¡ãŒã«ã¢ãã¬ã¹ãŸã§çŽæ¥ãé£çµ¡ãã ããã
ãææããã ãããããšãããããŸãããŸãããäžäŸ¿ããããããŸããããšããè©«ã³ç³ãäžããŸãã
æ¬å ·ã
[ãåå]
ã«ã¹ã¿ããŒãµãŒãã¹æ åœè ã
3.5 LoRA AdapterãããŒãžããŠåºåã確èª
AdapterãäœæãããŠããã®ã確èªã
!ls -l trained_t2l/mistral_7b_t2l/extras/user_generated/Adapterã®ããŒãžã«ã¯PeftãçšããŸãã
from peft import PeftModel
ADAPTER_DIR = "trained_t2l/mistral_7b_t2l/extras/user_generated/20250629_110411_CVNRpYo0/" # âã§è¡šç€ºããããã©ã«ãã«çœ®ãæã
lora_model = PeftModel.from_pretrained(base_model, ADAPTER_DIR)
lora_model = lora_model.merge_and_unload() # LoRA ãããŒã¹ã¢ãã«ã«ããŒãž
lora_answer = generate(lora_model, PROMPT)
print("â LoRA é©çšåŸ:\n", lora_answer)çµæã¯ä»¥äžã®ããã«ãªããŸãããTASK_DESCïŒã¿ã¹ã¯èª¬æïŒã«èšèŒããéãã察å¿ã¯ç°¡æœïŒæå€§2ïœ3æïŒãã§ãšããã®ãå®ãããŠããŸãã
ããŒãããœã³ã³ããŸã å±ããŠããªããšã®ããšããå°ãã§ãããããã®ä»¶ã«ã€ããŠèª¿æ»ããã§ããã ãæ©ããé£çµ¡ããããŸãã
4. ãŸãšã
ä»åã¯Text-to-LoRAãç°¡åã«è©ŠããŠã¿ãŸããã
Adapterã®ããŒãžã§ã¢ãã«ãå£ããã«ãã¡ããšèª¬æéãã®åºåãåºãããã«ãªã£ãã®ã¯ããããªãšæããŸããã
äžæ¹ã§ã
çŸç¶ã§ã¯æ¥æ¬èªã®ã¿ã¹ã¯ã«ã¯è»œã詊ããæãã§ã¯å¹ããªãã£ã
ããããLoRAã®å¿ èŠæ§ã¯ããã®ãïŒã¢ãã«æ§èœã®åäžã«ããã7Bãªã©ã®å°ããã¢ãã«ã§ãããã³ããã§ããçšåºŠå¶åŸ¡å¯èœïŒ
ãªã©ãè²ã
ãã£ãŠããããšã¯ãããªãšæããŸãããããŒã¿ã®æ©å¯æ§ãé«ãéèãè¡æ¿ã§ã¯LLM APIã®å©çšã«ããŒãã«ãããããšãå€ã
ããã®ã§ãããã£ããšããã«äœ¿ãããããããŸããã
ä»åã®text-to-loraã®ãããã¯ãŒã¯ã¯ç ç©¶ãšããŠããã€ãã®ãã³ãããŒã¯ã§åŠç¿ãããŠãããããå€ãã®ã¿ã¹ã¯ïŒæ¥æ¬èªå«ãïŒã§äœã£ãŠã¿ããã©ããªãããªã©æ°ã«ãªããŸãã
以äžãšãªããŸããè°è«ãææ³ãããã°NoteãX(Twitter)ã§ã³ã¡ã³ãã»ãæç€ºããã ãããšå©ãããŸãã
X: https://twitter.com/CurveWeb
ç®ãéããŠããã ãããããšãããããŸããã