tamdiep106/alpaca_lora_ja_en_emb-7b

This model is finetune on Japanese and English prompt

Usage:

Init model:

To use in code:

import torch
import peft
from transformers import LlamaTokenizer, LlamaForCausalLM, GenerationConfig

tokenizer = LlamaTokenizer.from_pretrained(
    "decapoda-research/llama-7b-hf"
    )

model = LlamaForCausalLM.from_pretrained(
    "tamdiep106/alpaca_lora_ja_en_emb-7b",
    load_in_8bit=False,
    device_map="auto",
    torch_dtype=torch.float16
    )

tokenizer.pad_token_id = 0  # unk. we want this to be different from the eos token
tokenizer.bos_token_id = 1
tokenizer.eos_token_id = 2

Try this model

To try out this model, use this colab space GOOGLE COLAB LINK

Recommend Generation parameters:

• temperature: 0.5~0.7

• top p: 0.65~1.0

• top k: 30~50

• repeat penalty: 1.03~1.17

Japanese prompt:

instruction_input_JP = 'あなたはアシスタントです。以下に、タスクを説明する指示と、さらなるコンテキストを提供する入力を組み合わせます。 リクエストを適切に完了するレスポンスを作成します。'
instruction_no_input_JP = 'あなたはアシスタントです。以下はタスクを説明する指示です。 リクエストを適切に完了するレスポンスを作成します。'

prompt = """{}
### Instruction:
{}

### Response:"""

if input=='':
    prompt = prompt.format(
        instruction_no_input_JP, instruction
        )
else:
    prompt = prompt.format("{}\n\n### input:\n{}""").format(
        instruction_input_JP, instruction, input
        )

result:

English prompt:

instruction_input_EN = 'You are an Assistant, below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.'
instruction_no_input_EN = 'You are an Assistant, below is an instruction that describes a task. Write a response that appropriately completes the request.'

prompt = """{}
### Instruction:
{}

### Response:"""

instruction = "write an email for my boss letting him know that i will resign from the position" #@param {type:"string"}
input = "" #@param {type:"string"}

if input=='':
    prompt = prompt.format(
        instruction_no_input_EN, instruction
        )
else:
    prompt = prompt.format("{}\n\n### input:\n{}""").format(
        instruction_input_EN, instruction, input
        )

result:

Use this code to decode output of model

for s in generation_output.sequences:
    result = tokenizer.decode(s).strip()
    result = result.replace(prompt, '')
    result = result.replace("<s>", "")
    result = result.replace("</s>", "")
    if result=='':
        print('No output')
        print(prompt)
        print(result)
        continue
    print('\nResponse: ')

    print(result)

Training:

Dataset:

• Jumtra/oasst1_ja

• Jumtra/jglue_jsquads_with_input

• Jumtra/dolly_oast_jglue_ja

• Aruno/guanaco_jp

• yahma/alpaca-cleaned

• databricks/databricks-dolly-15k

with about 750k entries, 2k entries used for evaluate process

Training setup

I trained this model on an instance from vast.ai

• 1 NVIDIA RTX 4090

• 90 GB Storage

• Time spend about 3 and a half days

• use python export.py to merge weight

• Training script: https://github.com/Tamminhdiep97/alpaca-lora_finetune/tree/master

Result

• Training loss

• Eval loss chart

Acknowledgement

• Special thank to KBlueLeaf and the repo https://huggingface.co/KBlueLeaf/guanaco-7b-leh-v2 that helped and inspired me to train this model, without this help, i wouldn't never thought that i could finetune a llm myself