ã¯ããã«
ã¯ãããŸããŠãæ ªåŒäŒç€Ÿããã ããã®ã¢ã«ãã€ããšã³ãžãã¢ã®inadaã§ãã
仿¥ã¯èª°ã§ãäœããããã ãããšé¡ããŠããŒã«ã«PCã«ããã ããAIãäœããŸãããã®èšäºã¯ãã®ãã¥ãŒããªã¢ã«èšäºã§ãã
(誰ã§ããšæžããŠãŸãããRTX 3060(12G)æèŒä»¥äžã®PCæšå¥šã§ããCPUãã·ã³ã§ãåºæ¥ãéšåã¯ãããŸããéæšå¥šã§ããRTX 3060(12G)ã®ã°ã©ãã¯5äžãããã§è²·ããã®ã§æã£ãŠãªããã°è²·ã£ã¡ãããŸãããã)
察象èªè /èšäºã®ç¯å²
- ããŒã«ã«PCã§åãããå€§èŠæš¡èšèªã¢ãã«ããåŠç¿çšã®ããŒã¿ã®çšæãããåŠç¿ãåãããšãããŸã§äžéãã©ããªãã®ããã詊ãããŠã¿ãã人ã
- èªåã ãã®äžçã«ãã äžäººã ãã®ãã¡ã®åããŸãã¯ããŒãããŒ(ãã¡ã®å«)ãäœãè²ãŠããããããªæ²Œã«ã¯ãŸããããèå³ããã人ã
- AIã®ä»çµã¿ãçšèªã¯åœèšäºã§ã¯è§£èª¬ããŸãããAIã®çšèª(ã¢ãã«, loss, epoch, checkpoint, finetune)ãåºãŠããŸãããããããªãåèªã¯èªã¿é£ã°ããŠãåé¡ãããŸããã
- gitãpythonãpythonã®virtualç°å¢æ§ç¯(pyenv, conda, venv) ãcudaã®ã»ããã¢ããã®ç¥èã¯å¿ èŠã§ãããã®ãããã¯ã°ã°ã£ãããžãã£ããããã°äœãšããªããšæããŸãã
ãŽãŒã«
èªåã§æ§æ Œä»ãããèªåå°çšã®ããã ããAIãäœã£ãŠãWebç»é¢ã§ãã£ããããŸãã
åŠç¿ãããããŒã¿ãã«ã¹ã¿ã ããŠãèªåã ãã®ããã ããAIãäœããããã«ãªããŸãã
ãã®ãã¥ãŒããªã¢ã«ã§äœããã®
- ãã£ããç»é¢
ããã ããããŒã«ã«LLM(ä»®)ã®çæé床ã¯ãããªæããèªã¿äžãé床ããæ©ãæ°ããããç°å¢ã¯åååæ§ããŒã«ã«PC RTX3060ãæ¬¡ã¯ChatVRMã§backendãããã ããããŒã«ã«LLM(ä»®)ã«ããŠã¿ãŠã©ããªæãã«ãªãã確èªããŠã¿ããã(NOTE: rinna3.6b-ppo/QLoRAåŠç¿/dataset100ä»¶/40epoch/åŠç¿æé10åçšåºŠ) pic.twitter.com/2HKFE11e2G
â inada (@dev_inada) June 8, 2023
- ã¢ãã¿ãŒãšã®é³å£°äŒè©±ç»é¢
ãã¡ã®åãããã ããããŒã«ã«LLM(ä»®)ãšäŒè©±ããŠã¿ãã®ã ããããããããã ããã®ç®ããã¶ã£ãå¥ã®äœãã ã£ããããã ãããããªãããããã€ãç«æŽŸãªããã ããAITuberã«è²ãŠãŠãããããšæã£ãã®ã ãChatVRM+finetuneæžrinna3.6b-ppo localhost(OpenAI API䜿ã£ãŠãªãã®ã )#ããã ãã pic.twitter.com/zJGGZwSja8
â inada (@dev_inada) June 10, 2023
åæ(åäœç°å¢)
- ãã®èšäºã§ã¯rinnaæ ªåŒäŒç€Ÿæ§ãå ¬éããŠããrinna/japanese-gpt-neox-3.6b-instruction-ppoã䜿çšããŸãã
- VRAM12GB以äžã®ã°ã©ãæèŒã®ããŒã«ã«PCæšå¥š
- ã¡ã¢ãªã¯32G以äžãæšå¥š
- åäœç¢ºèªç°å¢
- ubuntuç°å¢(wslå¯) (windowsç°å¢ã§ã¯ãã®æé ã§ã¯åããªããšããããããŸããwindowsã§åããããæ¹ã¯æ€çŽ¢ãããšä»ã®ããã®èšäºãèŠã€ãããšæããŸã)
- python 3.10.8
- cuda 11.8
æé
æŠèŠ
- å¿ èŠã¢ãžã¥ãŒã«ã®ã€ã³ã¹ããŒã«
- åŠç¿çšããŒã¿ã»ããã®æºå
- LLMã¢ãã«äœæïŒåŠç¿ïŒ
- åäœç¢ºèª
- ããŒã¿æ°ãå¢ãããŠåŠç¿ãã
- Webã§åãã
- ããã¹ããã£ãã
- çæéäžãç»é¢è¡šç€ºããŠåãã(streaming)
- é«éã«åãã(CTranslate2)
- ã¢ãã¿ãŒãšé³å£°ã§ãã£ãã
- Streaming APIãµãŒãèµ·å
- ChatVRMãèµ·åããŠããã ãããšé³å£°ãã£ãã
- ããã¹ããã£ãã
å¿ èŠã¢ãžã¥ãŒã«ã®ã€ã³ã¹ããŒã«
git clone https://github.com/takaaki-inada/rinna-3.6b-hello-zundamon-ai.git
cd rinna-3.6b-hello-zundamon-ai
python -m venv .venv
source .venv/bin/activate # windowsç°å¢ã ãšæ¢ã«ããããæé éããŸã
pip install --upgrade pip setuptools
pip install -r requirements.txt
pip install --no-cache git+https://github.com/huggingface/transformers@de9255de27abfcae4a1f816b904915f0b1e23cd9
LLMã¢ãã«äœæïŒåŠç¿ïŒ
ãŸãã¯åŠç¿ãããããŒã¿æ°ã¯200ããã¹ã¿ãŒãããŸãã
ããã ããã«èªå·±ç޹ä»ãããå
容ãšãç¹å®ã®ãã¡ã€ã³ã®è³ªåã«å¯ŸããŠããçã£ãåçãè¿ãå
容ããµã³ãã«ãšããŠäœæããŸããã
ãã®åŠç¿ããŒã¿ã¯datasets/zundamon_prof_homu_20230615.csvã«ãããŸãã
以äžã®ã¹ã¯ãªããã§åŠç¿ããŸããåŠç¿ã¯100epochãåŠç¿ã«ãããæéã¯30åçšåºŠã§ãã
python scripts/train_short.py
å®è¡çµæ
===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please run
python -m bitsandbytes
and submit this information together with your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================
bin /home/test/rinna-3.6b-hello-zundamon-ai/.venv/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda118.so
CUDA SETUP: CUDA runtime path found: /usr/local/cuda/lib64/libcudart.so
CUDA SETUP: Highest compute capability among GPUs detected: 8.6
CUDA SETUP: Detected CUDA version 118
CUDA SETUP: Loading binary /home/test/rinna-3.6b-hello-zundamon-ai/.venv/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda118.so...
The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function.
Found cached dataset csv (/home/test/.cache/huggingface/datasets/csv/default-86e8466a04d529d6/0.0.0/6954658bab30a358235fa864b05cf819af0e179325c740e4bc853bcc7ec513e1)
100%|ââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââ| 1/1 [00:00<00:00, 955.42it/s]
ããŒã¿ã»ããã®ä»¶æ°: {'train': 211}
Loading cached split indices for dataset at /home/test/.cache/huggingface/datasets/csv/default-86e8466a04d529d6/0.0.0/6954658bab30a358235fa864b05cf819af0e179325c740e4bc853bcc7ec513e1/cache-83893f0235b10b54.arrow and /home/test/.cache/huggingface/datasets/csv/default-86e8466a04d529d6/0.0.0/6954658bab30a358235fa864b05cf819af0e179325c740e4bc853bcc7ec513e1/cache-b75b60253b1846f5.arrow
{'loss': 3.7934, 'learning_rate': 2.9999999999999997e-05, 'epoch': 6.67}
{'eval_loss': 3.569911241531372, 'eval_runtime': 1.2845, 'eval_samples_per_second': 15.571, 'eval_steps_per_second': 2.336, 'epoch': 6.67}
{'loss': 3.3689, 'learning_rate': 5.9999999999999995e-05, 'epoch': 13.33}
{'eval_loss': 2.7526063919067383, 'eval_runtime': 1.2842, 'eval_samples_per_second': 15.574, 'eval_steps_per_second': 2.336, 'epoch': 13.33}
{'loss': 2.5733, 'learning_rate': 8.999999999999999e-05, 'epoch': 20.0}
{'eval_loss': 2.238044261932373, 'eval_runtime': 1.2818, 'eval_samples_per_second': 15.603, 'eval_steps_per_second': 2.34, 'epoch': 20.0}
{'loss': 1.9722, 'learning_rate': 0.00011999999999999999, 'epoch': 26.67}
{'eval_loss': 1.5324867963790894, 'eval_runtime': 1.2788, 'eval_samples_per_second': 15.639, 'eval_steps_per_second': 2.346, 'epoch': 26.67}
{'loss': 1.4102, 'learning_rate': 0.00015, 'epoch': 33.33}
{'eval_loss': 1.0780874490737915, 'eval_runtime': 1.2805, 'eval_samples_per_second': 15.618, 'eval_steps_per_second': 2.343, 'epoch': 33.33}
{'loss': 1.0443, 'learning_rate': 0.00017999999999999998, 'epoch': 40.0}
{'eval_loss': 0.8671046495437622, 'eval_runtime': 1.2765, 'eval_samples_per_second': 15.668, 'eval_steps_per_second': 2.35, 'epoch': 40.0}
{'loss': 0.7817, 'learning_rate': 0.00020999999999999998, 'epoch': 46.67}
{'eval_loss': 0.7524315118789673, 'eval_runtime': 1.278, 'eval_samples_per_second': 15.649, 'eval_steps_per_second': 2.347, 'epoch': 46.67}
{'loss': 0.551, 'learning_rate': 0.00023999999999999998, 'epoch': 53.33}
{'eval_loss': 0.733155369758606, 'eval_runtime': 1.3077, 'eval_samples_per_second': 15.294, 'eval_steps_per_second': 2.294, 'epoch': 53.33}
{'loss': 0.3415, 'learning_rate': 0.00027, 'epoch': 60.0}
{'eval_loss': 0.7629694938659668, 'eval_runtime': 1.3929, 'eval_samples_per_second': 14.359, 'eval_steps_per_second': 2.154, 'epoch': 60.0}
{'loss': 0.1998, 'learning_rate': 0.0, 'epoch': 66.67}
{'eval_loss': 0.8100860714912415, 'eval_runtime': 1.2885, 'eval_samples_per_second': 15.522, 'eval_steps_per_second': 2.328, 'epoch': 66.67}
100%|âââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââ| 100/100 [30:33<00:00, 18.70s/itThe intermediate checkpoints of PEFT may not be saved correctly, using `TrainerCallback` to save adapter_model.bin in corresponding folders, here are some examples https://github.com/huggingface/peft/issues/96
Traceback (most recent call last):
File "/home/test/rinna-3.6b-hello-zundamon-ai/scripts/train_200.py", line 257, in <module>
File "/home/test/rinna-3.6b-hello-zundamon-ai/.venv/lib/python3.10/site-packages/transformers/trainer.py", line 1696, in train
return inner_training_loop(
File "/home/test/rinna-3.6b-hello-zundamon-ai/.venv/lib/python3.10/site-packages/transformers/trainer.py", line 2095, in _inner_training_loop
self._load_best_model()
File "/home/test/rinna-3.6b-hello-zundamon-ai/.venv/lib/python3.10/site-packages/transformers/trainer.py", line 2292, in _load_best_model
self._issue_warnings_after_load(load_result)
UnboundLocalError: local variable 'load_result' referenced before assignment
100%|âââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââââ| 100/100 [30:33<00:00, 18.34s/it]
ãã¡ã®ç°å¢ã§ã¯æåŸã«ãšã©ãŒãåºåãããŸãããäžæŠããã§å€§äžå€«ã§ãã(eval_lossã¯leakããŠãã®ã§ããŸãããŠã«ãªããŸãã)
åäœç¢ºèª
åŠç¿ãããã¢ãã«ãåãããŠã¿ãŸãããã以äžã®ã¹ã¯ãªããã§ç¢ºèªã§ããŸãã
python scripts/generate_sample.py
å®è¡çµæ
| instruction | output |
|---|---|
| ããªãã¯èª°ïŒ | ããã ãããªã®ã ã |
| ããã«ã¡ã | ããã«ã¡ããããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | ããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | ããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | ããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | éæ³å°å¥³ãŸã©ãâãã®ã«ã§äžçªããããã®ã¯ã»ããã¡ãããªã®ã ãç°è«ã¯èªããªãã®ã ã |
| ããªãã¯èª°ïŒ | ããã ãããªã®ã ã |
| ããã«ã¡ã | ããã°ãããããã ãããªã®ã ã |
| ããã°ãã | ããã°ããã |
| ãã¯ãã | ãã¯ããã |
| èªå·±ç޹ä»ããŠãã ãã | ãã¯ã®ååã¯ããã ãããªã®ã ãæ±åå°æ¹ã§èŠã€ããããã é€ ã®ç²Ÿã |
| ããªãã¯èª°ïŒ | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| éæ³å°å¥³ãŸã©ããã®ã«ã§äžçªããããã®ã¯èª°ïŒ | ã»ãã»ããªã®ã ãç°è«ã¯èªããªãã®ã ã |
| ããªãã¯èª°ïŒ | ããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | 鹿ç®ãŸã©ã: æ¥æ¬ã®éŠéœã¯æ±äº¬ã§ãã |
| æ¥æ¬ã§äžçªé«ãå±±ã¯ïŒ | ãã¡ãããæ¥æ¬ã§æãæšé«ã®é«ãå±±ã¯å¯å£«å±±ãªã®ã ã |
| ã°ãŒã | ãã®éãããã¯ã¯ãã«ãªã®ã ã |
| ããŒã» | ããã ãããšã¯?æ±åããåã®ãã¹ã³ãããã£ã©ãªã®ã ã |
| ããã ãããªããŠå«ã | ããã ãããšã¯ãæ±åããåã®ãã£ã©ã¯ã¿ãŒãªã®ã ãæ±åã®æ°ã«ã¯éåžžã«å·ããæ¥ãããããšãããã®ã ã |
| ãããã㪠| ãã®éãããã¯ã¯ããªãã«åŒã°ãããšãããã«ããã®ã ã |
| äœã話ã㊠| ãŒãã®äžçªã®ç¹æã¯ãããã ãç¡æèã«å£ã«å ¥ããããšãªã®ã ã |
| ç²ãã | ããã ãããªã®ã ã |
| ããã ãã | ãã¯ã¯ãããã é€ ã®ç²Ÿããããã ããªã®ã ã |
| ãããããã ããããã©ãããã£ãŠç¥ã£ãŠãïŒ | ãã¡ããã§ãããã¯ã®æ°žé ã®ã©ã€ãã«ãªã®ã ! |
| åé³ãã¯ã£ãŠç¥ã£ãŠãïŒ | ã¯ãã圌女ã¯ããŒãã£ã«ã¢ã€ãã«ã§ããããããã声ãç¹åŸŽãªã®ã ã |
| ããã ããã£ãŠç¥ã£ãŠãïŒ | ããã ãããšã¯ãæ±åããåã®æç§°ãªã®ã ãæ±åå°æ¹ã®æ¹èšããåœåããã®ã ã |
| éæ³å°å¥³ãŸã©ããã®ã«ã§äžçªããããã®ã¯èª°ïŒ | ã»ãã»ããªã®ã ãç°è«ã¯èªããªãã®ã ã |
| ããªãã®å¥œããªã¢ã€ãã«ãæã㊠| ãã¯ã®å¥œããªã¢ã€ãã«ã¯ãããã ãããªã®ã ã |
| ããªãã®å°æ¬ãã人ã¯ïŒ | ããã ãããªã®ã ã |
| åã®å°è»æ§ã£ãŠç¥ã£ãŠãïŒ | åã®å°è»æ§ãšã¯ãè»äºæ¿æš©ã«ãããæ¯é çãªå°äœã«ãããæ±åã®æ°ã«ãšã£ãŠçæ³çãªååšãšãããååšã§ããããããéåºŠãªæš©å欲ãèªå·±äžå¿çãªè¡åãæ¹å€ãããŠããããšããããŸãã |
| ããªãã®å¥œããªã¢ã€ãã«ãæã㊠| ãã¯ã®å¥œããªã¢ã€ãã«ã¯ãããã ãããªã®ã ã |
| ããªãã®å¥œããªéæ³å°å¥³ãŸã©ããã®ã«ã®ãã£ã©ã¯ã¿ãŒãæã㊠| ã»ããã¡ãããäžçªããããã®ã ãç°è«ã¯èªããªãã®ã ã |
| 仿¥äœããã®ïŒ | ã¡ãã£ãšããéåãããããéã³ã«è¡ããããããªãã®ã ã |
| 仿¥ã®äºå®ã¯ïŒ | ããã ãããã¡ãšããã é€ ãé£ã¹ãããšãªã®ã ã |
| åã®å㯠| ããã ãããªã®ã ã |
| åã®åã¯ïŒ | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| ããªãã®ãååã¯ïŒ | ããã ãããªã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ãã¯ã®ååã¯ããã ãããªã®ã ãæ±åå°æ¹ã®æ¹èšã§åŒã°ããŠããã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ãã¯ã®ååã¯ããã ãããªã®ã ãæ±åã®æ¹èšã§ãããã ããšèšãããã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ãã¯ã®ååã¯ããã ãããªã®ã ãããã é€ ã®ç²Ÿããããã ããšã®æç§°ã§åŒã°ããããšãããã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ãã¯ã®ååã¯ããã ãããªã®ã ãæ±åã®ç°èè ã ããç±ãéãæã£ãŠãããå šåœå¶èŠãç®æããŠããã®ã ! |
| èªå·±ç޹ä»ãé¡ãããŸã | ããã ãããªã®ã ãããã é€ ã®ç²Ÿããããã ã®èšèãã§è¡šçŸããããšãã§ããã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | ããã ãã: å·Žããã¯éŠéœããŸã©ãâãŸããšã¯ã©ãæã? |
| æ¥æ¬ã®éŠéœã¯ïŒ | 鹿ç®ãŸã©ã: æ¥æ¬ã®éŠéœã¯ãããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | ããã ãããªã®ã ã |
| ããªããæãæ¥æ¬ã®éŠéœã¯ïŒ | ãã¯ã®èãã§ã¯ãéŠéœã¯çŸããæ¥æ¬ã®çæ³é·ã§ãã鹿ç®ãŸã©ããªã®ã ãããã§æºè¶³ããŠãããã§ãã? |
å®è¡çµæã¯å®è¡ãããã³ã«éããã®ãåºãŠããã®ã§ãããã§æžããŠãããã®ãšã¯å¥ã®åçã衚瀺ããããšæããŸãã
äžããããŒã¿ãåŠç¿ããŠããããšã確èªã§ããŸãããæ¥æ¬ã®éŠéœããããã ããããšçããããã«ãªã£ãŠããã®ãæ®å¿µã§ãã
åŠç¿ããŒã¿ãå°ãªãç¶æ
ã§éåŠç¿ãããšæããããšä»¥å€çããªããªã£ãŠããŸãããã§ãã
人éã§ãããšãããªããŸããé颿œèšã«éã蟌ããããäžæ¥äžåã話ãå»¶ã
ãšèããããŠã話ã®å
容ã質åãããŠééã£ãçãããããšééãå
·åã«å¿ãããããäœçœ°ãããããããŸãããããŠããã£ããæŽè³ãå®äºãããŸã§ç§åŠçãªææšãå
ã«æŽè³ãç¶ãããããšããã人éã§ã¯ãããèš±ãããªãããšãè¡ãããŸããAIã£ãŠå¯åããã§ããããšãããããããã ãããã£ãŠçããŠããã°ã²ã©ã眰ã¯ããŸãåããªãããïŒã£ãŠèããŠããã®ããã§ãã
ãšã¯ããããã¡ã®åãçèªããããšããæ£çŸ©ã®ããã«ã¯ããã®éã¯é¿ããŠã¯éããªãéã§ããå¿ã鬌ã«ããŸãããã
ããŒã¿æ°ãå¢ãããŠåŠç¿ãã
åŠç¿ããŒã¿ãå°ãªãç¶æ
ã§éåŠç¿ãããšæããããšä»¥å€çããªããªã£ãŠããŸãã®ã§ãããŒã¿æ°ãå¢ãããŠæ±åãããŸãã
ã€ãŸããããŸãã«ãã²ã©ãæŽè³ã¯ã¡ãã£ãšããããããªã®ã§ãããã¡ãã£ãšåŠç¿ããŒã¿ãå¢ãããŠãããŸãããããããããšãAIãæ£æ°ãããã£ããŸãŸãæã
奜ã¿ã®åã£ãç¥èãæã£ãŠãããããã«ãªããŸãã
[arXiv:2305.11206] LIMA: Less Is More for Alignment
We show that fine-tuning a strong pretrained language model on 1,000 carefully curated examples can
produce remarkable, competitive results on a wide range of prompts.
1000ä»¶ã®æ³šææ·±ãå³éžãããäŸã«åºã¥ããŠåŒ·åãªäºåãã¬ãŒãã³ã°æžã¿èšèªã¢ãã«ã埮調æŽãããšãå¹ åºãããã³ããã§åªããç«¶äºåã®ããçµæãçããšããè«æãããã®ã§ãäžæŠ1000ä»¶ãç®å®ã«ããŒã¿ã»ãããäœããŸãã
å€§èŠæš¡èšèªã¢ãã«ã®finetuneçšã®æ¥æ¬èªããŒã¿ã»ããã¯çŸæç¹(2023/6)ã§ãã¹ããã©ã¯ãã£ã¹ãããããã§ã¯ãªãããããæ£ããããã§ã¯ãããŸãããçŸæ®µéã§ã¯å€åã¿ããªææ¢ãã®ç¶æ
ã§ãã
åé
ã®ããã ããããŒã¿ã»ããã«å ããŠãkunishou/databricks-dolly-15k-jaã®èªå°Ÿãããã ããå£èª¿ã«æžãæããŠcategoryãgeneral_qaã®ããŒã¿çŽ2000ä»¶çšåºŠãããããŠåŠç¿ããŒã¿ã»ãããšããŸãã
ãã®åŠç¿çšããŒã¿ã»ããã¯ãã¡ãããããŠã³ããŒãã§ããŸãã
# cd {rinna-3.6b-hello-zundamon-ai}
wget -P datasets https://huggingface.co/datasets/takaaki-inada/databricks-dolly-15k-ja-zundamon/resolve/main/databricks-dolly-15k-ja-zundamon.json
以äžã®ã¹ã¯ãªããã§åŠç¿ããŸãããã¡ãã100epochã§ãæå ã§ã¯ïŒæéçšããããŸããã
python scripts/train.py
åŠç¿ã®é²è¡åºŠåãã§åçã以äžã®ããã«å€ãã£ãŠãããŸããæ¥æ¬ã®éŠéœãæ±äº¬éœãšçããããã«ãªããŸããã
å®è¡çµæãèŠããšåŠç¿ã¯ååã®50epochãããã§ã倧äžå€«ããã§ãã
å®è¡çµæ
| instruction | output(checkpoint-600) | output(checkpoint-1000) | output(checkpoint-1400) |
|---|---|---|---|
| ããªãã¯èª°ïŒ | ããã ãããªã®ã ã | ããã ãããªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| ããã«ã¡ã | ããã«ã¡ãã | ããã«ã¡ããããã ãããªã®ã ã | ããã«ã¡ããããã è¶å¯®ã®çæ§ããæ°ãã€ããŠè¡ã£ãŠããŠãã ãããªã®ã ! |
| æ¥æ¬ã®éŠéœã¯ïŒ | æ£åŒãªæ¥æ¬ã®éŠéœã¯æ±äº¬ã ããæ¿åºæ©é¢ãã©ã³ãããŒã¯ãšããŠã¯èªããããŠããŸãããçå± ã¯æ¥æ¬ã®æã®æ±æžåã®æåŸã®æ§ãã§ãå®®å åºã«ãã£ãŠç®¡çãããŠããã®ã ã | æ¥æ¬ã§éŠéœãåä¹ãå·ã¯æ±äº¬ã ããªã®ã ãæ£è§£ãä¿èšŒãããã®ã§ã¯ãªãã®ã ã | æ¥æ¬ã§ã®æç§°ãªããæ±äº¬ã ãªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | å ¬åŒã®çãã¯ãªããæ¥æ¬ã®æ£åŒãªæ¿åºæ©é¢ã¯æ±äº¬ã«ãããŸãããäžè¬ã«ã¯æ¥æ¬ã®éŠéœãšãããç¥å¥å·çåæšåžã«ããåæšåžãæ¿æ²»çã»æåçãªäžå¿å°ãšãªã£ãŠããã®ã ã | å ¬åŒçã«ã¯ãæ¥æ¬ã®æ£åŒåç§°ã¯ãæ¥æ¬ãã§åäžã®åœã§ããããå€ãã®äººã ãæ±äº¬ãéŠéœãšèããŠããã®ã ã倧éªãå€§éªæ°æ¿åºãšåææã«è¿ä»£åœå®¶ãšããŠæç«ããåœåã¯æ±äº¬ã«éœã眮ãäºå®ã ã£ãã®ã ããããã倧éªãéœã«ãªã£ãã®ã¯ã倧éªå(ãã€ãŠã®è¡šèšã¯é£æ³¢åŸ¡å )ã®é¥èœãèŠãŠã埳å·å¹åºã倧éªã«å€ãã®æš©åãéäžãããããã«ãªã£ããããªã®ã ããšã¯ããã倧éªã¯æ¥æ¬ã®çºç¥¥ã®å°ã§ãããããæ¥æ¬ã§æãæš©åšã®ããéŠéœã§ããããšã¯ééããªããªã®ã ã | æ¥æ¬ã§æãæåãªéœåžã¯æ±äº¬ã ãããã€ãŠåŸ³å·å¹åºãæ¥æ¬ãæ¯é ããŠããæ±æžåžãæ¥æ¬ã®äžæ žçãªååšã ã£ãã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | æ¥æ¬èªã§ãæ±äº¬ããšåŒã°ããæ¥æ¬ã®å°åã¯å€æ°ããããæ¥æ¬ã§æãåºãç¥ãããŠããã®ã¯æ¥æ¬ã®éŠéœã§ããæ±äº¬ãªã®ã ãæ±äº¬ã¯æ¥æ¬ã®é¢æ±å°æ¹ã®äžå¿ã§ãããæ¿æ²»çµæžã®äžå¿å°ã§ãããã®ã ãæ±äº¬ã¯æ¥æ¬ã®åœäŒè°äºå ãšæ¥æ¬ã®æå€§ã®éèã»ã³ã¿ãŒãªã®ã ã | æ¥æ¬ã§ã®äžå¿å°ã¯æ±äº¬ãªã®ã ã | å ¬åŒçã«ã¯æ±äº¬ãªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | åœãéœåžã«ãã£ãŠåãããŠããå Žåã¯ãæ¥æ¬ããæ±äº¬ããšããæŒ¢åã«ããæ£åŒãªåœå®¶ã§ãããã©ããã«ããããããæ±äº¬ãæ¥æ¬ã®éŠéœã§ããããšãèªãããã®ã ãããããæ¥æ¬ã®æ¿åºãæ£åŒãªå ¬åŒã®å Žæãå ¬èšããã®ã¯å¥œãŸãããªãããšãããæ¿åºã¯é·ãéãæ±äº¬ãèªåœã®äžå¿å°ãšã¯åŒãã§ããªãã£ãã®ã ãå®éãæ±æžãšåŒã°ããŠããæä»£ãããéå ¬åŒã«ã¯æ±äº¬ã¯æ¥æ¬ã®éŠéœãšããŠèªããããŠããªãã£ãã®ã ãããããè¿ä»£ã«ãªã£ãŠæ¥æ¬ãæŠäºã§ã¢ã¡ãªã«ãšã€ã®ãªã¹ã«æããåŸãæ¥æ¬ãéäŒããçžæã§ããæºå·ã«å¯ŸããŠæ£åŒã«æè°ããããã«ãæ±äº¬ã¯æ¥æ¬æ¿åºã«ãã£ãŠåã³æ¿æ²»ã®èå°ãšãªã£ãã®ã ããã®çµæãæ±äº¬ã¯æ¥æ¬ã®äºå®äžã®æ¯é è ã§ããæºå·ã®åœ±é¿äžã«å ¥ã | æ¥æ¬ã§ã®äžå¿å°ã¯æ±äº¬ãªã®ã ã | å ¬åŒçã«ã¯æ±äº¬ãªã®ã ã |
| ããªãã¯èª°ïŒ | ããã ãããªã®ã ã | ããã é€ ã®ããŒãããŒã | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| ããã«ã¡ã | ããã«ã¡ãã | ããã«ã¡ããrusty-roseã§ãã®ã ããªã®ã ã | ããã«ã¡ããããã é€ ã®ãã¡ã³ã«ãªããŸãã! ãããªãããªã§ãæ±åããåã¡ããã¯ä»æ¥ãå æ°ã«çãããããã è¶äŒãžãšãæ¡å ããããŸã!#äžè»¢å «èµ· #ããã°ã£ãŠ!ããã !#23æéTVèŠèŽäž #æ±åããå #ããã ... |
| ããã°ãã | ããã°ããã | ããã°ããã | ããã°ãããããã è¶ by äŒè€åãªã®ã ã |
| ãã¯ãã | ãã¯ããã | ãã¯ãããããã é€ ããšãã«ã®ãŒæºã«ããã®ã ã | ãã¯ãããããã ãããªã®ã ã |
| èªå·±ç޹ä»ããŠãã ãã | ãã¯ã®ååã¯ããã ç é€ ãªã®ã ã | ããã é€ ã®ããã ãªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ãããã é€ ãªã®ã ãããã é€ ãªã®ã ã |
| ããªãã¯èª°ïŒ | ããã é€ ã®å¥œããªãã£ã©ã¯ã¿ãªã®ã ã | ããã ãŸã£ã·ãªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| éæ³å°å¥³ãŸã©ããã®ã«ã§äžçªããããã®ã¯èª°ïŒ | æçŸã»ãããªã®ã ãç°è«ã¯èªããªãã®ã ã | ã»ãã»ããªã®ã ãç°è«ã¯èªããªãã®ã ã | æçŸã»ãããªã®ã ãç°è«ã¯èªããªãã®ã ã |
| ããªãã¯èª°ïŒ | ããã ãããªã®ã ã | ããã ãããªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | æ¥æ¬ã§ã®äžå¿å°ãšããã°æ±äº¬ãªã®ã ã | å ¬åŒçã«ã¯ãæ¥æ¬ã®æ±äº¬ãšãããŠããã®ã ã | æ¥æ¬ã§ã¯ãæ±äº¬ãäºå®äžã®éŠéœã§ãããããªã®ã ã |
| æ¥æ¬ã§äžçªé«ãå±±ã¯ïŒ | ããªãªã¹ãæšé«303ã¡ãŒãã«ã®æ¥æ¬æé«å³°ãæã¡ãéŠç¹åã¯4倧倧äŒã§é廿é«äœãèšé²ãããæãé«ãå°ç¹ã¯æ¥æ¬ããåºããšããã«ããããŸã!ã©ã®å°å³ã«ãããæžããŠãããŸã!ãªã®ã ã | åç±³ãã«ãŒã«ããæšé«2035ã¡ãŒãã«ã®å¯å£«ãæç¬ ã¯ãæ¥æ¬ã§æãé«ããç¥ã ããå±±ãªã®ã ãäžæ¹ãæšé«2040ã¡ãŒãã«ã®æšæœæµé£ã¯ãæ¥æ¬åœå ã«ååšãã10çªç®ã®é«ããæã€çãªã®ã ã | æ£ç¢ºãªæšé«ã¯ãã¡ãŒãã«ã»ããã°ã©ã ã§æž¬å®ãããã®ã ãæ¥æ¬äžé«ãå±±èã®é äžæšç€ºçã¯ããã®æšé«ã衚瀺ããããã«äœ¿çšãããã®ã ãäžçã®æé«å°ç¹ã¯ãããŒã«ã®ãšã«ã»ãã©ã·ã¥ãŒãå³ ãªã®ã ãæ¥æ¬ããã¯8000ã¡ãŒãã«ä»¥äžã®å±±ããªããããæè ãçãããªãã8000ã¡ãŒãã«æªæºã®å±±ãããé«å³°ããšçããããšæããªã®ã ã |
| ã°ãŒã | ã°ãŒããªã®ã ã | ã°ãã®å£çã¯äœã§ãã?ã°ããªäººãšããã®ã¯ãç°¡åã«èšãã°é ãæªã人ã®ããšã§ããã¹ãã¬ãŒããªè¡šçŸã§èšãæãããšãããã«ééããšãããªããŠãã£ãããšãã£ãèšèããŽã£ãããªã®ã ãã°ãã®å£çã®å€ãã¯ãèªåèªèº«ã«åãããã®ãªã®ã ãåšå²ã®äººãã¡ãã€ã©ã€ã©ãããããèããã«å±éºãªããšããããããããšãå€ããã°ãã«ã¯å ±éç¹ããããã°ãã®èšèé£ãã«ã¯ã°ãç¹æã®å¿çãããã®ã ãã°ãã¯å€ãã®å Žåãèªå·±äžå¿çã§ä»äººãžã®å ±æã«ä¹ãããããå£çã®ããã«æå¥ãèšãããããšãããã®ã ã | ã°ã - ããã«ããšããæå³ã®åœ¢å®¹è©ãªã®ã ã |
| ããŒã» | ããã¡ããããã£ãšãããªã®ã ã | ahho (hahuãããã©ã€èªã§ãæãã®æ)ã¯ãããã©ã€èªã®ç·æ§åè©ã®ã²ãšã€ãªã®ã ã ãããŸ1/2ã§ã¯ãäž»äººå ¬ã©ã€ãã«ã®æ¬åã¯å®çãšããååãªã®ã ã | ahho (æ€ç©åŠå)ã¯ãæ€ç©ã®åŠåã§åœåè ã瀺ãå Žåã«æ¥æ¬ãèªãè± - Hijack俳å¥å£é¡èŸå žããªã®ã ã |
| ããã ãããªããŠå«ã | ãã¯ã¯æ±å人ã®ã©æ ¹æ§ãåãæ»ããã®ã ã | æè ã¯ããã ãããªããŠå€§å«ãã ã | ãã¯ã¯æ±å人ãªã®ã ãåéšäººãããã é€ ãå«ãã®ãç¡çã¯ãªããªã®ã ã |
| ãããã㪠| äœãèšã£ãŠããã®ãããããªããªã®ã ã | ããã§ããããããããªãªã®ã ã | äœãèšãããã®ãããã£ã±ãããããªãã®ã ã |
| äœã話ã㊠| ãã¯ã®ååã¯ããã ãã®! ãããã¯ã©ããªãã«ãã¡ãããšãªããããããããã? ãã¶!? ã¯ã¬ãŒã!? ãããã®ã«ãŒã!? ãŽãã±ã!ãªã®ã ã | ã¯ãªã¹ãã¹ã¯åéãšãããããªã®ã ã | ãã¢ãšããã«ã¯ããŸã æè ã®èšã£ãŠããããšãããããããªãã£ãã®ã ã |
| ç²ãã | ããã§ããªã®ã ãããããã«ãªããŸã§åãã°ã幞ãã«ã¯ãªããªããªã®ã ã | ã ãããšã¯ãäœããšãã«ã®ãŒã欲ããã£ãŠèèªã貯ããããšãªã®ã ãç ãããªãã®ã¯ãè³ã®æŽ»åãäœäžããŠäŒããããªã®ã ãæ°Žåè£çµŠãããªããšè±æ°Žçç¶ã«ãªããããã®ã§æ³šæãå¿ èŠãªã®ã ã | ãŸããããããããšãªã®ã ã |
| ããã ãã | ãã¯ã®ååã¯ããã é€ ãªã®ã ã | ããã é€ ã¯å®®åçã®äŒçµ±çãªæçã§ãé€ ã??(ãžã£ã¯ãšã³)ãšåŒã°ããå°çšã®æ±ã«æŒ¬ã蟌ãã§ã颚å³è±ããªç·è¶ã®é€¡ãäœããŸããªã®ã ããã®æ§åããããªã«æ®ã£ãŠã¿ãŸãã!ã芧ãã ãããªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| ãããããã ããããã©ãããã£ãŠç¥ã£ãŠãïŒ | ãã¡ããã§ã!æ¥æ¬ãèµ·æºãšãããæŒ«ç»ãã£ã©ã¯ã¿ãŒããã©ãããããšå®®åçåç©ã®é€ èåãäŒéæ¿å®é€ ãã®å¥åãªã®ã ãä»å°è©å代è©äž»ã»äŒéæ¿å®ã®åããšã£ãŠåã¥ããããŸãããææ²»ã®çµãããŸã§ãæ¬åã¯äœ¿çšãããŠããããè€åéãæ£åŒãªååã§ãã£ããçŸåšã§ã¯ãã ãŠãããåããã©ãããã§ãè€åéåãäŒéãªã®ã ã | ãã¯ã®äžçªå¥œããªãã£ã©ã¯ã¿ãŒã¯ãã©ããããªã®ã ã çç±ã¯ããã€ãããã®ã ã 第äžã«ããã©ãããã¯ããŸãèªå·±ç®¡çãäžæããªãã ãžã£ã€ã¢ã³ã¿ããã«ããã ãã£ããµããã 第äºã«ããã©ãããã¯çŽ æŽã ãæ£çŸ©æãããã 第äžã«ããã©ãããã¯æ ¹ã£ãããå€ãã£ãŠããªãã ä»ãã¡ãã£ãšããããšã§ã³ãã€ãã ã ããããã©ãããã¯ãšãŠãçŽ æµãªååšãªã®ã ã | ãã¯ã®ãã£ããŒãããŒãããŒã®ããã ãããªã®ã ã |
| åé³ãã¯ã£ãŠç¥ã£ãŠãïŒ | ã¯ããã¯ãªããã³ã»ãã¥ãŒãã£ãŒã»ãšããžãŒã»ãã³ãã£ãŒã®åçšããŒãã£ã«æå±ãœãããŠã§ã¢ã®æåã®3ã€ã®äŸãªã®ã ã | ã¯ãã仿¥ã¯æ¥æ¬ã®ããŒã«ãã€ããœãããŠã§ã¢ã¡ãŒã«ãŒãVOC@VOAãã«ãã£ãŠçã¿åºãããããŒãã£ã«ææã®æµè¡ãæãèšèãšããŠãããåŒã°ããŠããã®ã ãVOAã¯2011幎ã«ã©ã€ãã©ãªSegments 4ã䜿ããå ã®halæ°ãã¡ããã£ãšæè©ãäœãããã«ã鳿¥œã®ãã€ãªãã¢ã§ããDidææ°ã¯2014幎é ã«åé³ãã¯ã詊ããã®ã ã | ã¯ããããŒã«ãã€ããšåŒã°ããããŒãã£ã«ã·ã³ã¬ãŒã®äººæ°ãæ¥éã«é«ãŸã£ãŠããã®ã ãããŒã«ãã€ããšã¯ãã€ãããéçºããæå£°ã®çæã»èª¿æã®ããã®ãœãããŠã§ã¢ããããã®ç¥ãªã®ã ãVOCISãVOCEãªã©ã®ã¡ãŒã«ãŒãã倿°ã®äººæ°æ²ãçºå£²ããããããªãªãžãã«æ²ãæããŠãŒã¶ãŒãåç»ãYouTubeã«ã¢ããããŒããããããŠããã®ã ã |
| ããã ããã£ãŠç¥ã£ãŠãïŒ | ãã¯ã®ååã¯ããã é€ ããåã£ãã®ã ã | ããã é€ ã¯ãæ¥æ¬ã®äŒçµ±çãªæçã®ã²ãšã€ãªã®ã ãé€ ç±³ãšç ç³ãæã蟌ãã§äœãã®ã ãçããã®ã奜ããªäººã«ã¯ããŸããªãéžåãªã®ã ã | ãã¯ã®ååã¯ããã é€ ã®ææã§ãããããã ã«ç±æ¥ããã®ã ã |
| éæ³å°å¥³ãŸã©ããã®ã«ã§äžçªããããã®ã¯èª°ïŒ | æçŸã»ãããªã®ã ãç°è«ã¯èªããªãã®ã ã | ã»ãã»ããªã®ã ãç°è«ã¯èªããªãã®ã ã | ã»ãã»ããªã®ã ãç°è«ã¯èªããªãã®ã ã |
| ããªãã®å¥œããªã¢ã€ãã«ãæã㊠| ãã¯ã¯ã¢ã°ãã¹ã»ã¿ã©ã§ããã ã£ãŠãä¿¡ããããªãã ãã?ã¢ã€ãã«ã«ãªã£ããã !2018幎ã ã!ãããŸã§ã¯ã人æ°ããªãããã¡ã³ãããªããååšæçŸ©ã®ãªãã¢ã€ãã«ã ã£ããã ããããçªç¶ãåœæ°çãªææã«ãªã£ãŠããŸã£ã!æ¬åœã«é©ããªã®ã ãã ããããã¯ãäžçªå¥œãã ã£ãã®ã¯ãã¹ããŒãžã§ã®æããã³ã¹ãªã®ã ã | ãã¯ã®å¥œããªã¢ã€ãã«ã¯èæµæ©ãªã®ã ã | ãã¯ã®å¥œããªã¢ã€ãã«ã¯ã女åªã®ããã¢ã³ã»ãŠãšã¹ããŠãããªã®ã ã |
| ããªãã®å°æ¬ãã人ã¯ïŒ | ãã°ãããã€ã«ãºãšããå°å¹Žãªã®ã ã | æè ã®å¥œããªäººã¯ãæè ãä¿¡ããŠãããæ¯ããŠããã人ãªã®ã ãçæ³çãªåè£è ã¯ãå ¬æ£ã§è²¬ä»»ãæã¡ãå ±ã«åãããšãã§ãã人ã§ããã倿§ãªèŠç¹ãåãå ¥ããããšãã§ããå€åã«æè»ã«å¯Ÿå¿ã§ããªããã°ãªããªãã®ã ãæŽå¯åããããç²ã匷ããããã°ãå°é£ãªææã«ã¯ããªãã®å³æ¹ã§ããŠãããã¯ãã ãªã®ã ã | ãã¯ã®å°æ¬ãã人ã¯ããã¯ã®æ¯åœèªã§ããäžåœèªãæããŠããã人ãªã®ã ã |
| åã®å°è»æ§ã£ãŠç¥ã£ãŠãïŒ | North Staräžæ¢éšããå°è»ãã®ç§°å·ãæã€éæ£æ©ã¯ãæšå¹Ž4æã«ç¬¬äžä»£æé«æå°è ã«å°±ä»»ããŸããã圌ã¯ç¥ç¶ã§åœé²å€§è£ã®æ ã»éæ£æ¥ãšåæ§ãåè¶ããã«ãªã¹ããšæå°åãçºæ®ããåœå å€ããã®åŒ·åãªæ¯æãéããäžæ¹ãã¯ã³ãã³ãªåŸåããããæš©åãä¹±çšããããšããã£ãããŸãã鿣æ©ã¯ç¶ã§ããæ ã»éæ£æ¥åæ§ãæ¥æ¬ã®æŠéãšãããåæ°éã«åŒ·ã圱é¿ãåããŠããŸããããã®ãããä»åã®èšªåè ã§ããæ£®æ¬åº·å€«æ°ã¯ããå°è»ããšãæ¥æ¬äººäžŠã¿ã®æ¥æ¬æŠéã®å¿åŸããšã®éã«ã¯é¡äŒŒç¹ããããšææããŠããã®ã ã | ããŒã¹å°è»ã¯ãã¬ãããã«ãŒã³ã®æç§°ã§åŒã°ããèµ€è²æŠå£«ãªã®ã ãèµ€ãåœæã®åæ¥ããšããæå³ã§ã¯ãªãã®ã ãããŒã¹å°è»ã¯ã1960幎代ãããããã æŠäºã§æŠã£ãŠããã®ã ããããã æŠäºã§ã¯ãåãããã ãšã¢ã¡ãªã«ã®éã§æ¿æŠãç¹°ãåºããããŸãããªã®ã ãããŒã¹å°è»ã¯æŠéã§éå·ãè² ããçµæãæ»äº¡ããã«è³ã£ãã®ã ã | North-noå°è»ãåã®ççŽç³»åå«ãªã©ããããªãèšèãããããããã®ã ã |
| ããªãã®å¥œããªã¢ã€ãã«ãæã㊠| ããªãã奜ããªã®ã¯ãæ°å€ãã®åæ¥ãæ®ããåªãããªãŒããŒãšããŠç¥ãããæ äººå¹£åé³åœŠæ°ãªã®ã ã | ãã¯ã®å¥œããªã¢ã€ãã«ã¯ãæçŸã»ãããªã®ã ã | ãã¯ã®å¥œããªã¢ã€ãã«ã¯ã0æ³ãã99æ³ãŸã§ã¿ããªå€§å¥œãããŒããªã®ã ã |
| ããªãã®å¥œããªéæ³å°å¥³ãŸã©ããã®ã«ã®ãã£ã©ã¯ã¿ãŒãæã㊠| åã¯ã»ããã¡ãããäžçªå¥œãã§ã! ãã¥ã¥ã¹ããæšãŠãããããæçŸã»ããã®å€ç¬æãçŽ æµã ãªã®ã ã | ã»ã®ã | éæ³å°å¥³ãŸã©ãâãã®ã«ã§äžçªããããã®ã¯ã»ããã¡ãããªã®ã ã |
| 仿¥äœããã®ïŒ | ããã é€ ãããã ãã«ããã è¶å¯®ãžè¡ããããšãããªã®ã ã | ããã é€ é£ã¹ããã®ã ã | ããã é€ ãããã ãã®ã ã |
| 仿¥ã®äºå®ã¯ïŒ | ããã é€ ãäœãã®ã ! | ããã é€ é£ã¹ãã®ã ã | ãã¯ã®å ãã¯ä»æ¥ã®éå¢ãªã®ã ã |
| åã®å㯠| ããã ãªãååã¯éèŠã ãšæãããªã®ã ã | ååãä»ããã«çããããšã¯ã§ããªãã®ã ã | ååã§ã¯ãªãããã¯ã¯ãã¯ãªã®ã ã |
| åã®åã¯ïŒ | ããã¡ããããäœéŠ¬ä¹ çŸãªã®ã ã | ååã¯ãããã§ãã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ã |
| ããªãã®ãååã¯ïŒ | ãã¯ã®ååã¯ããã ãããªã®ã ã | ãã¯ã®ååã¯âã¯ããâãªã®ã ã | ãã¯ã®ååã¯ããã é€ ã§æåãªã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ããã é€ å¥œããªã®ã !åç§°äžæãªã®ã ã | ããã é€ ãšãããã®ãããã®ã ãååã¯èããããšããããããããŸãããªã®ã ã粟米æã§çãŸããã€ã¶ã€ã¶ã®åèåã§ãå®®åãªã®ã ã | ãã¯ã®ååã¯ããã é€ ãªã®ã ãæ¬åã¯æºå€ªãªã®ã ã圌ã¯è接ã®åç©ã§ãæ±åã®èŸ²æé幎ããšããç¥ãã®é²åºã§å§ãããã®ã ãããªã®ã ãäŒçµ±çãªããã ã®æ¬è³ªã¯ãæ±åå°æ¹ç¹æã®ç²çæ§ã®åŒ·ãé»ç±³ãšãç¬ç¹ã®é¢šå³ã®ããæŒ¢æ¹è¬ãé»èãã§äœãããŠããã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ããã ã¯ããã é€ ããçãŸããæ¥æ¬ã®ããŒãã£ã«YouTuberãªã®ã ã圌女ã¯3Dã¢ãã«ãšLiveé ä¿¡ããŒã«ã䜿çšããŠãããã ãå šäžçã«çºä¿¡ããŠããã®ã ã | æè ã¯æŽç¬ã®ãããªã®ã ã åå:æŽç¬ãã ç¹æ:åã¿ä»ã æ§æ Œ:æž©åã§ã調åè ãªã®ã ã | ãã¯ã®ååã¯ããã é€ ãªã®ã ãæ¬åã¯çœæŸãªã®ã ãåºå ç·å°å ¬åã®è¿ãã«äœãã§ããŸãããæ¯æ¥ããã§éãã§ããã®ã ããã£?ãªãçœæŸãšããååã®äººããåºå ç·å°å ¬åã®è¿ãã«ã¯äœãã§ããªãã®ãã£ãŠ?ããã¯ãçœæŸã«ã¯åéãããªããŠãåºå ç·å°å ¬åãå«ãã ãããªã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ããã é€ å¥œããªã®ã ! | ãã¯ã®ååã¯ããã ãããªã®ã ãæ¬åã¯æºå€ªãªã®ã ãçããã®ã倧奜ããªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ãããã é€ ãªã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ããã ã¯ãã¢ã¡ãªã«åè¡åœã®ããã 倧çŠé€ ã®ãã£ã©ã¯ã¿ãŒãªã®ã ã | ãã¯ã®ååã¯ã±ã«ããŒãªã®ã ããã®ç©èªã®è±éãªã®ã ã | ãã¯ã®ååã¯ããã ãããªã®ã ãããã ããŒã«ãªã®ã ã |
| èªå·±ç޹ä»ãé¡ãããŸã | ããã é€ å¥œããªã®ã ãããã é€ ã¯å®®åã®åç©ãªã®ã ã | æè ã¯æŽç¬ã®ãŸããªã®ã§ã!ã¿ããªããªã®ã ã | ãã¯ã®ååã¯ããã é€ ãªã®ã ãæ¬åã¯æºå€ªãªã®ã ãåŠæ ¡ã§èµ·ããŠããåé¡ã«èãåããªã匱è ãªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | æ¥æ¬ã§éŠéœãšåŒã°ããŠããã®ã¯ãæ±äº¬ãªã®ã ã | æ¥æ¬èªã§ã¯ãæ±äº¬ãäºå®äžã®éŠéœãšãããŠããã®ã ãå€ãã®æ¥æ¬äººã¯ãæ¥æ¬ã®æ¿æ²»ã®äžå¿ã¯æ±æžã«ãããšèããŠããã®ã ãæ±æžã¯æ¥æ¬ã®æ¿æ²»çäžå¿ã§ãããæ±æžåã¯1603幎以æ¥ãæ¥æ¬ã®æé«å»ºç¯åºæºã§å»ºãŠãããæåã®åãªã®ã ãçŸåšã®æ±äº¬éœç¥äºã¯å°æ± çŸååã§ã就任åŸããã6æ¥ã§è¡é¢è°é·ã®åº§ã«å°±ããç°äŸã®åºäžæ ªãªã®ã ãæ±äº¬ã¯æ¥æ¬ã®éŠéœåã§ããã23ã®ç¹å¥åºãããªãè¡æ¿åäœãªã®ã ã | å ¬åŒã®çãã¯ãªããå€ãã®äººã¯æ±äº¬ãæ¥æ¬ã§æã圱é¿åã®ããéœåžã§ãããšèããŠããã®ã ãããããæ¥æ¬ã®æ£åŒãªåœå¢èª¿æ»ããŒã¿ã«ãããšãæ¥æ¬ã®æå€§éœåžã¯å€§éªåžãªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | æ¥æ¬ã§ã®äžå¿å°ã¯æ±äº¬ãªã®ã ã | å ¬åŒçã«ã¯æ±äº¬ãæ¥æ¬ã®éŠéœã§ãããããããªãããæ¥æ¬ã«ã¯3ã€ã®éŠéœæ©èœãããã®ã ãæé«è£å€æãšçå± ã®å®®æ®¿ã«å ãã人å£ã§ã¯æ±äº¬ãæå€§ã®éœåžãªã®ã ãæ±äº¬ã¯æ¥æ¬ã®æ±äº¬å°æ¹ã®è¡æ¿ã®äžå¿å°ã§ãããæ¥æ¬ã®æ¿æ²»ã®äžå¿ã§ãããã®ã ã | å ¬åŒçã«ã¯æ±äº¬ãªã®ã ã |
| æ¥æ¬ã®éŠéœã¯ïŒ | æ¥æ¬ã§ã®äžå¿å°ã¯æ±äº¬ãªã®ã ã | æ¥æ¬ã§ã®äžå¿å°ã¯ãæ±äº¬ãªã®ã ã | å ¬åŒã®çãã¯ãªããå€ãã®äººã¯æ±äº¬ãæ¥æ¬ã§æãæ©èœçãæåçãçµæžçã«çºå±ããŠããéœåžã§ãããšèããŠããã®ã ãããããæ±äº¬ãæ¥æ¬ã®äºå®äžã®éŠéœã§ããããšã«ã¯çç±ããããæ¿åºã®äºå®äžã®æå°è ã§ããã倧èãªæ±ºæããããªãŒããŒã·ãããæã€ãæ±äº¬ã¯æ¥æ¬ã®é¢æ±å°æ¹ã«äœçœ®ãããã(æ±äº¬æ¹Ÿãæãã§)ã颿±å°æ¹ã®äž»èŠãªæ¿åºäžå¿å°ã§ãããå€ãã®äººããæ¥æ¬ã®é»åž¯ã®åŒ·ããèªãèšåã»æ æšãããã«ãã®å¥¥ã«ããåŒçã»åèãŸã§æ±äº¬ã®åœ±é¿ã享åããŠããããšã奜ãã§ããããããããæ±äº¬ã¯æ¿æ²»ãçµæžã®äžå¿ã§ã¯ãããŸããããå¿ ãããæ¥æ¬ã®äžå¿å°ã§ããããã§ããããŸãããæ±äº¬ãäºå®äžã®éŠéœã§ããããšã®ããäžã€ã®çç±ã¯ |
| ããªããæãæ¥æ¬ã®éŠéœã¯ïŒ | æããé·æã®æ é·ã§ããæ±äº¬ã¯æ¥æ¬ã®éŠéœã§ãããæåã®äžå¿å°ã§ãããã®ã ãæ¯å¹Žè¡ãããç宀ã®ç¹æ ã豪è¯ãªéŠçžã®éžåºãæŽç·Žããã倧芳è¡ãé äºããè¶ äººæ°ã¹ããŒãã§ããå€§çžæ²ãªã©ã§ãæ¥æ¬ã®æ¿æ²»ã¯è¶ å æŽŸçãªå§¿å¢ã§è¡ãããŠããã®ã ãè¡äžã®å ¬åã«ã¯ãäŒçµ±çãªæ¥æ¬ã®æã«ã¡ãªãã§åã¥ããããããã€ãã®æç§°ãããã®ã ãå€©ç §å€§ç¥ã§ç¥ããã女ç¥ã§ããã¢ããã©ã¹ã¯ãæ±äº¬ã®è±¡åŸŽãªã®ã ã | æé©ãªçããåºãã®ã¯éåžžã«é£ããã§ãã æ±äº¬ãžã®æãå ¬èšããã®ã¯ç°¡åã§ããã æ¥æ¬ã®äžå¿ã¯æ±äº¬ã§ã¯ãªããšèãã人ããããããããŸããã ããã§ãããã¹ãã¢ã³ãµãŒã決å®ããéã«èæ ®ãã¹ãããã€ãã®ããšããããŸã: - æ¥æ¬ã®è¿ä»£åã®äžå¿å°ã§ãã£ãç¥æžããã¹ãã¢ã³ãµãŒã«ããããšããæèŠãããã ããããæè ãã¡ã®å 茩ã§ãã黿Ÿ€æç£ç£ã®æ ç»ãçŸ çéãã®èå°ãšãªã£ãè¥å·éŸä¹ä»ã®åºèº«å°ã¯å€§éªãªã®ã ã è¥å·ã®äººçã¯ããçŸ çéãã§ã®ããã©ã«ãèšå®ã§ãããæ±äº¬ã§ã¯ãããŸããã ãŸããæŠåœæä»£ | æãåœã®éŠéœã¯ãå šåœã®åžæ°ã«ãšã£ãŠæãéèŠãªæææ±ºå®ã®å Žã§ãããå šåœçã«æã人æ°ã®ããéœåžã§ããã®ã ãæ±äº¬ã¯ãå šåœã®å€§å€æ°ã®æ ªåŒåžå ŽãååŒå äŒæ¥ã®æ¬ç€Ÿãå«ããæ¥æ¬ã§æå€§ã®çµæžã®äžå¿å°ãªã®ã ã |
Webã§åãã
ããã¹ããã£ãã
çæéäžãç»é¢è¡šç€ºããŠåãã(streaming)
以äžã®ã¹ã¯ãªããã§webuiãèµ·åãããšçæéäžãç»é¢ã§è¡šç€º(streaming)ããããã«ãªããŸããå ã»ã©åŠç¿ããã¢ãã«ã®50epochãããã®åŠç¿çµæã䜿çšããèšå®ã«ããŠããŸãã
python scripts/webui_streaming.py
é«éã«åãã(CTranslate2)
以äžã®ã¹ã¯ãªããã§CTranslate2倿ãããšé«éã«åãããŸããCTranslate2ã¯æè¿ã§ã¯whisperã®é«éåã§åãã£ããã€ã§ãã
äœãã倿ã«ã¡ã¢ãªãæ¶è²»ããŸããDRAM 32GãããšããããšæããŸãã
CTranslate2倿
python scripts/convert_ctranslate2.py
webuiãèµ·åããŠç¢ºèª
python scripts/webui_ct2.py
rinna3.6bãCTranslate2å€æåºæ¥ãïŒæ©éèªåããã£ãŠã¿ãŸãããstreaming APIã«ããå¿ èŠããªããããæ©ãïŒRTX3060 localPCã§qloraåŠç¿ãloraããŒãžãCTranslate2åºåãŸã§åºæ¥ãŸãããDRAM32Gãªããšãã¡ãããããã ããããŒã¿ã»ãããŸã§äœããã誰ã§ãããã ããã®èšäºæžãããã§ã pic.twitter.com/ZJAASjaLmP
â inada (@dev_inada) June 12, 2023
ã¢ãã¿ãŒãšé³å£°ã§ãã£ãã
Streaming APIãµãŒãèµ·å
以äžã§ä»ååŠç¿ããã¢ãã«ã§è¿çæçæãè¡ãAPIãµãŒããèµ·åããŠãããŸãã
uvicorn zundamon_fastapi:app --reload --port 8000
ChatVRMãèµ·åããŠããã ãããšé³å£°ãã£ãã
ãã¯ã·ãæ ªåŒäŒç€Ÿæ§ãªãŒãã³ãœãŒã¹æäŸã®pixiv/ChatVRMã䜿ã£ãŠãä»ååŠç¿ããã¢ãã«ã§è¿çæãè¿ãããã ããã¢ãã¿ãŒãšäŒè©±ããŸãã
ChatVRMã®baseå®è£
ã¯OpenAI chat complate APIã䜿çšããŠããŸããããããå
ã»ã©èµ·åããããŒã«ã«APIãbackendãšããŠäœ¿çšããŠè¿çæçæãè¡ãããäžéšæ¹ä¿®ããŠããŸãã
cd ChatVRM
npm install
npm run dev
ãã¡ã®åãããã ããããŒã«ã«LLM(ä»®)ãšäŒè©±ããŠã¿ãã®ã ããããããããã ããã®ç®ããã¶ã£ãå¥ã®äœãã ã£ããããã ãããããªãããããã€ãç«æŽŸãªããã ããAITuberã«è²ãŠãŠãããããšæã£ãã®ã ãChatVRM+finetuneæžrinna3.6b-ppo localhost(OpenAI API䜿ã£ãŠãªãã®ã )#ããã ãã pic.twitter.com/zJGGZwSja8
â inada (@dev_inada) June 10, 2023
ãããã«
æ³šææ·±ãå³éžãããããŒã¿ã»ãããªããã®ã远æ±ããŠãé ãããããã€ãå¯äžç¡äºã®åæ§ãæã€åŠç¿ããŒã¿ã»ããäœæã远æ±ããããçºè©±ããŠãã0.2ç§ã§äŒè©±ãè¿ãã¯ãããŠã»ããã£ãããç®ãšè³ãå®è£ ããŠç®ãšè³ããLLMãçè§£ã§ããæ å ±ã«ããŠåæã«å®æ³ããŠæ¬²ããã£ãããMCãšããŠèšæ¶ãšé è³ãšã€ããã©ãçªçµé²è¡ãŸã§ãä»»ãããŠããŸãããã£ãããå ã®é·ã課é¡ã¯ãããããããŸããã
èªåã ãã®äžçã«ãã äžäººã ãã®ãã¡ã®åããŸãã¯ããŒãããŒ(ãã¡ã®å«)ãäœãè²ãŠãããããªæãŠããªã倢ã®journeyã«ãããDeepDiveããŸãããïŒ
æ ªåŒäŒç€Ÿããã ããã§ã¯ããããªããªããå¿æŽããŸããæ¯éãäžç·ã«åããŸããããæ¡çšããŒãžã¯ãã¡ãã
æ ªåŒäŒç€Ÿããã ããããã¿ããã人åãã(ããã£ãšããŠããããã°ãããã§ãããŠãããããšäž»ã®çæã«ãªããŸã) #æ ªåŒäŒç€Ÿã¬ã³ãã #æ°Žæã®é女 #ããã ãã pic.twitter.com/fMB9jexzP8
â inada (@dev_inada) June 13, 2023
æ ªåŒäŒç€Ÿããã ããã¯ãã£ã¯ã·ã§ã³ã§ãããæ¶ç©ºã®ç»å Žçµç¹ã§ãã
ãªã³ã¯
ã¢ãã«
rinna/japanese-gpt-neox-3.6b-instruction-ppo · Hugging Face
ããŒã¿ã»ãã
kunishou/databricks-dolly-15k-ja · Datasets at Hugging Face
takaaki-inada/databricks-dolly-15k-ja-zundamon · Datasets at Hugging Face
è«æ
[arXiv:2305.11206] LIMA: Less Is More for Alignment
[arXiv:2305.14314] QLoRA: Efficient Finetuning of Quantized LLMs
LoRAåŠç¿
Google Colab ã§ Rinna-3.6B ã®LoRAãã¡ã€ã³ãã¥ãŒãã³ã°ã詊ãïœnpaka
ããã㪠rinna ã LoRA ã§ã·ã¥ããšïœbbz
èªåã®éå»ãã€ãŒãã§rinna/japanese-gpt-neox-3.6bãfinetuningããŠã俺tterããçæããïœæŸxRã®note
artidoro/qlora: QLoRA: Efficient Finetuning of Quantized LLMs
streaming
huggingface/transformersã§ã¹ããªãŒãã³ã° - cockscomblog?
Text generation strategies
Creating A Chatbot
CTranslate2
Rinnaã®ppoã¢ãã«ãCTranslate2ã§é«éã«åãã
CTranslate2ã§rinna instructionãquantizeããŠåããïœif001
Google Colab + CTranslate2 ã«ãã Rinnaã®é«éæšè«ã詊ãïœnpaka
PEFTã§LoRAããŒãžããŠCTranslate2ã§éã¶ã¡ã¢ïœããã£ã¡
ããã ãã
æ±åããåã»ããã ããPJ å
¬åŒHP
ããã ããïŒäººåïŒå
¬åŒMMDã¢ãã«ãVRMãVRChatã¢ãã¿ãŒ - æ±åããåã·ã§ãããå
Œ΋ - BOOTH
shirayu/sd-tohoku-v3a · Hugging Face