metadata
base_model: tomo1222/gemma-2-27b-bf16-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- gemma2
- trl
license: gemma
language:
- jp
datasets:
- llm-jp/magpie-sft-v1.0
- tomo1222/Japanese-QA111dataset
Uploaded model
- Developed by: tomo1222
- License: Gemma
- Finetuned from model : tomo1222/gemma-2-27b-bf16-4bit
tomo1222/gemma-2-27b-bf16-4bit : google/gemma-2-27bをUnslothで直接用いるために、BitsAndBytesを用いて4bit量子化し、そのまま保存したもの。
This gemma2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
output code
library
pip install unsloth
pip install --no-deps --upgrade "flash-attn>=2.6.3"
pip install -U ragatouille
pip install fugashi unidic-lite
inference code using Google Colaboratory(L4)
from datasets import concatenate_datasets, load_dataset
from unsloth import FastLanguageModel
import random
import json
from huggingface_hub import login
from google.colab import userdata
login(userdata.get('HFtoken'))
with open("elyza-tasks-100-TV_0.jsonl","r",encoding='utf-8') as f:
tasks = [json.loads(l) for l in f.readlines()]
model_name = "tomo1222/Gemma2-27b-ft-jp-r64_alpha64"
max_seq_length = 4096
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = model_name,
max_seq_length = max_seq_length,
dtype = None,
load_in_4bit = True,
)
# google/gemma-2-9bのテンプレート
tokenizer.chat_template = """
{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}
"""
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
dataset = load_dataset("tomo1222/Japanese-QA111dataset")
ref_tasks = list(dataset["train"])
ref_tasks_input = [task["input"] for task in ref_tasks]
dic = {}
dic_input = {}
for i, task in enumerate(ref_tasks):
dic[ref_tasks_input[i]] = task["output"]
dic_input[ref_tasks_input[i]] = task["input"]
"""# 2. RAGのロード"""
from ragatouille import RAGPretrainedModel
RAG = RAGPretrainedModel.from_pretrained("bclavie/JaColBERTv2")
RAG.encode(ref_tasks_input)
def search_ref_input(input, k=10):
retreived=RAG.search_encoded_docs(query=input,k=k)
print(retreived)
text ="質問・文章をよく読んで、正確で親切な回答を書きなさい。\n"
for data in retreived[::-1]: # inverse order
key = data["content"]
output = dic[key]
input = dic_input[key]
text+="### 質問:\n"+input+"\n\n### 回答:\n"+output+"\n\n\n"
return text
"""# Prompt"""
output_data=[]
for i, task in enumerate(tasks):
text = (
search_ref_input(task["input"], 20)
+ "あなたは日本語が堪能な優秀な人間です。\n"
+ "**文脈**を踏まえて、改行と箇条書きを駆使して、日本語で**詳細に**書きなさい。\n"
+ "優秀な人間になりきって、推測をいれずに根拠をもってわかりやすく答えてください。"
+ f"### 質問:\n{task['input']}\n\n### 回答:\n"
)
print(task["input"])
inputs = tokenizer(text, return_tensors="pt").to("cuda")
print(len(inputs['input_ids'][0]))
output = model.generate(**inputs, max_new_tokens=1024,repetition_penalty=1.1,use_cache=True,
bad_words_ids = [tokenizer.encode("質問", add_special_tokens=False),
tokenizer.encode("###", add_special_tokens=False),
tokenizer.encode("#", add_special_tokens=False),
tokenizer.encode("##", add_special_tokens=False),
tokenizer.encode("---", add_special_tokens=False),
tokenizer.encode("<h3>", add_special_tokens=False),
tokenizer.encode("filepath", add_special_tokens=False),
tokenizer.encode("> ", add_special_tokens=False),
]
)
output_text = tokenizer.decode(output[0][inputs.input_ids.size(1):], skip_special_tokens=True).strip()
print(i,output_text)
print("---")
output_data.append({"task_id":i,"output":output_text})
with open("output.jsonl","w",encoding="utf-8") as f:
for result in output_data:
json.dump(result, f, ensure_ascii=False)
f.write('\n')