README.md

metadata

base_model:
  - oxygen65/llm-jp-3-13b-finetune-2
  - llm-jp/llm-jp-3-13b
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - llama
  - trl
license: cc-by-nc-sa-4.0
language:
  - ja
datasets:
  - elyza/ELYZA-tasks-100

How to Use

1. load this model and tokenizer

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
)
import torch
from tqdm import tqdm
import json

model_name = "oxygen65/llm-jp-3-13b-finetune-3"

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

# Load model
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

2. load Eval Datasets

tasks = []
with open("./elyza-tasks-100-TV_0.jsonl", "r") as f:
    item = ""
    for line in f:
      line = line.strip()
      item += line
      if item.endswith("}"):
        tasks.append(json.loads(item))
        item = ""

3. set up retrievers

if you can't find "rank_bm25" python package in your environment

!pip install rank_bm25

from rank_bm25 import BM25Okapi
from nltk.tokenize import word_tokenize
import nltk
import numpy as np


# 必要なデータをダウンロード（初回のみ）
nltk.download('punkt')
nltk.download('punkt_tab')

def search_similar_documents_bm25(query, sample_tasks):
    # トークン化（BM25はトークン化されたデータを要求します）
    tokenized_documents = [word_tokenize(doc) for doc in sample_tasks['input']]

    # BM25オブジェクトの作成
    bm25 = BM25Okapi(tokenized_documents)

    tokenized_query = word_tokenize(query)
    # 類似度の計算
    doc_scores = bm25.get_scores(tokenized_query)
    # 類似度が高い順にソート
    sorted_indexes = np.argsort(doc_scores)[::-1]

    indexes = []
    for i in range(len(doc_scores)):
        if doc_scores[sorted_indexes[i]] < 20.0:
            break
        else:
            indexes.append(sorted_indexes[i])
    
    return indexes

from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
SentTF = SentenceTransformer('all-MiniLM-L6-v2')
def seearch_similar_documents_neuralRetriver(query, sample_tasks):
    global SentTF
    emb1 = SentTF.encode([query])
    emb2 = SentTF.encode(sample_tasks['input'])
    # 全ての組み合わせで類似度を計算
    similarity_matrix = cosine_similarity(emb1, emb2) #時間かかるので先に計算しておくべき
    # 類似度が高い順にソート
    sorted_indexes = np.argsort(similarity_matrix[0])[::-1]
    #print(sorted_indexes)
    
    indexes = []
    for i in range(len(sample_tasks['input'])):
        if similarity_matrix[0][sorted_indexes[i]] < 0.75:
            break
        else:
            indexes.append(sorted_indexes[i])
    
    return indexes

def create_icl_prompt(input, sample_tasks, task_id):
    indexes_bm25 = search_similar_documents_bm25(input, sample_tasks)
    indexes_neu = seearch_similar_documents_neuralRetriver(input, sample_tasks)
    indexes = list(set(indexes_bm25 + indexes_neu))
    icl_prompt = ""
    if indexes == []:
        return ""
    
    icl_prompt = f"""## 例題\n"""
    for i in range(len(indexes)):
        icl_prompt += f"""### 指示
{sample_tasks["input"][indexes[i]]}
### 回答
{sample_tasks["output"][indexes[i]]}
"""
    icl_prompt += f"""
## 本題: 以下の指示に従って回答してください。step by stepで回答してください。
"""
    return icl_prompt 

create_icl_prompt(tasks[2]["input"], sample_tasks, 0)

4. Inference

# llmjp
import re
pattern = r"^以下.*$"

# プロンプトの作成
sys_prompt = ""
icl_prompt = ""
results = []
loop = 0
for data in tqdm(tasks):
  task_id = data["task_id"]
  input = data["input"]
  # in context learning用のプロンプト
  icl_prompt = create_icl_prompt(input, sample_tasks, task_id)
  
  prompt = f"""{sys_prompt}{icl_prompt}### 指示
{input}
### 回答
"""  
  tokenized_input = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt").to(model.device)
  with torch.no_grad():
      outputs = model.generate(
          tokenized_input,
          max_new_tokens=512,
          do_sample=False,
          repetition_penalty=1.2,
          eos_token_id=tokenizer.eos_token_id,
      )[0]
  output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)

  while (True): #とりあえず出力。
    line = output.splitlines()
    if re.match(pattern, line[0]) and len(line) == 1:
      print(f"#=========================  Unexpected answer =========================#\n {line}")
      outputs = model.generate(
          tokenized_input,
          max_new_tokens=512,
          do_sample=True,
          temperature=0.4,
          repetition_penalty=1.2
      )[0]
      output = tokenizer.decode(outputs[tokenized_input.size(1):], skip_special_tokens=True)
    else: break  


  results.append({"task_id": data["task_id"], "input": input, "output": output})
  
  print(f"task_id: {data['task_id']}, prompt: {prompt}, output: {output}")
  

5. Dump results

import re
model_name = re.sub(".*/", "", model_name)
with open(f"./{model_name}-outputs.jsonl", 'w', encoding='utf-8') as f:
    for result in results:
        json.dump(result, f, ensure_ascii=False)  # ensure_ascii=False for handling non-ASCII characters
        f.write('\n')

Uploaded model

• Developed by: oxygen65

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