taoki's picture
Update README.md
86089b8 verified
|
raw
history blame
2.71 kB
metadata
language:
  - ja
license: apache-2.0
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - trl
  - mistral
datasets:
  - sakusakumura/databricks-dolly-15k-ja-scored
  - llm-jp/oasst1-21k-ja
  - nu-dialogue/jmultiwoz
  - kunishou/amenokaku-code-instruct
license_name: mistral
base_model: tokyotech-llm/Swallow-MS-7b-v0.1

Uploaded model

  • Developed by: taoki
  • License: apache-2.0
  • Finetuned from model : tokyotech-llm/Swallow-MS-7b-v0.1

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained(
    "taoki/Swallow-MS-7b-v0.1-qlora-oasst1-jmultiwoz-dolly-amenokaku"
)
model = AutoModelForCausalLM.from_pretrained(
    "taoki/Swallow-MS-7b-v0.1-qlora-oasst1-jmultiwoz-dolly-amenokaku"
)

if torch.cuda.is_available():
    model = model.to("cuda")

prompt="[INST] ไปŠๆ—ฅใฏ4/1ใชใฎใงใ™ใŒใ€ๅ‘จใ‚Šใฎ็š†ใ•ใ‚“ใŒ็ชๆ‹ๅญใ‚‚ใชใ„ใ“ใจใ‚’่จ€ใฃใฆใ„ใฆๅ›ฐๆƒ‘ใ—ใฆใ„ใพใ™ใ€‚ไธ€ไฝ“ไฝ•ใŒ่ตทใ“ใฃใฆใ„ใ‚‹ใฎใงใ—ใ‚‡ใ†ใ‹๏ผŸ [/INST]\n"

input_ids = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **input_ids,
    max_new_tokens=512,
    do_sample=True,
    top_p=0.95,
    temperature=0.1,
    repetition_penalty=1.1,
)
print(tokenizer.decode(outputs[0]))

Output

<s> [INST] ไปŠๆ—ฅใฏ4/1ใชใฎใงใ™ใŒใ€ๅ‘จใ‚Šใฎ็š†ใ•ใ‚“ใŒ็ชๆ‹ๅญใ‚‚ใชใ„ใ“ใจใ‚’่จ€ใฃใฆใ„ใฆๅ›ฐๆƒ‘ใ—ใฆใ„ใพใ™ใ€‚ไธ€ไฝ“ไฝ•ใŒ่ตทใ“ใฃใฆใ„ใ‚‹ใฎใงใ—ใ‚‡ใ†ใ‹๏ผŸ [/INST]
4ๆœˆ1ๆ—ฅใฏใ‚จใ‚คใƒ—ใƒชใƒซใƒ•ใƒผใƒซใงใ€ไบบใ€…ใฏๅ†—่ซ‡ใ‚„ใ„ใŸใšใ‚‰ใ‚’่จ€ใฃใฆๆฅฝใ—ใ‚€ๆ—ฅใจใ•ใ‚Œใฆใ„ใพใ™ใ€‚ใ“ใฎ็ฟ’ๆ…ฃใฏใ€1564ๅนดใซใƒ•ใƒฉใƒณใ‚นใฎใ‚ทใƒฃใƒซใƒซ9ไธ–ใŒ4ๆœˆ1ๆ—ฅใซ็ตๅฉšใ—ใŸใ“ใจใ‹ใ‚‰ๅง‹ใพใฃใŸใจ่จ€ใ‚ใ‚Œใฆใ„ใ‚‹ใ€‚

ใ—ใ‹ใ—ใ€ใ‚ใชใŸใŒๅ›ฐๆƒ‘ใ—ใฆใ„ใ‚‹ใฎใชใ‚‰ใ€ใใ‚ŒใฏใŠใใ‚‰ใใ€ใ‚ใชใŸใŒๅ†—่ซ‡ใ‚„ใ„ใŸใšใ‚‰ใ‚’่จ€ใฃใฆใ„ใ‚‹ไบบใŸใกใŒใ€ใ‚ใชใŸใŒใใฎใ‚ˆใ†ใชใ‚‚ใฎใ ใจๆ€ใฃใฆใ„ใชใ„ใ“ใจใ‚’่จ€ใฃใฆใ„ใ‚‹ใ‹ใ‚‰ใ ใจๆ€ใ„ใพใ™ใ€‚ใ“ใ‚Œใฏใ€ๅฝผใ‚‰ใŒใ‚ใชใŸใ‚’้ฉšใ‹ใ›ใŸใ‚Šใ€็ฌ‘ใ‚ใ›ใŸใ‚Šใ™ใ‚‹ใŸใ‚ใซๅ˜˜ใ‚’ใคใ„ใฆใ„ใ‚‹ๅฏ่ƒฝๆ€งใŒใ‚ใ‚‹ใ“ใจใ‚’ๆ„ๅ‘ณใ—ใพใ™ใ€‚

ใ‚‚ใ—่ชฐใ‹ใŒใ‚ใชใŸใ‚’ๅ›ฐๆƒ‘ใ•ใ›ใŸใ‚Šใ€ไธๅฟซใซใ•ใ›ใŸใ‚Šใ™ใ‚‹ใ‚ˆใ†ใชใ“ใจใ‚’่จ€ใฃใŸๅ ดๅˆใฏใ€ๆฐ—ใซใ›ใš็„ก่ฆ–ใ—ใฆใใ ใ•ใ„ใ€‚ใพใŸใ€่‡ชๅˆ†่‡ช่บซใ‚„ไป–ไบบใ‚’ๅ‚ทใคใ‘ใ‚‹ใ‚ˆใ†ใชๅ˜˜ใ‚’ใคใใ“ใจใฏ้ฟใ‘ใพใ—ใ‚‡ใ†ใ€‚</s>

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.