metadata

base_model: llm-jp/llm-jp-3-13b
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - llama
  - trl
license: apache-2.0
language:
  - en

Uploaded model

Developed by: ikedachin
License: apache-2.0
Finetuned from model : llm-jp/llm-jp-3-13b

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

使用したdataset

下記からランダムに5000データを抽出

DeL-TaiseiOzaki/Tengentoppa-sft-v1.0
llm-jp/magpie-sft-v1.0

実行コード

from tqdm import tqdm
import os
import json

import torch
from unsloth import FastLanguageModel

from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    BitsAndBytesConfig,
)

HF_TOKEN = "your-token"
model_name = "ikedachin/llm-jp-3-13b-ozaki-ds-5000"

# QLoRAの設定
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=False,

)

# modelのダウンロード
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    token = HF_TOKEN
)

# tokenizerのダウンロード
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, token = HF_TOKEN)


prompt = "<ここに入力を入れる>"

# トークン化
tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)

# 推論
with torch.no_grad():
    outputs = model.generate(
        tokenized_input,
        max_new_tokens=300,
        do_sample=False,
        repetition_penalty=1.2
    )[0]

# トークンから言葉にデコード
output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)