---
library_name: transformers
license: cc-by-4.0
datasets:
- DeL-TaiseiOzaki/Tengentoppa-sft-v1.0
- llm-jp/magpie-sft-v1.0
- ntotsuka123/clean3-ultraboros-20k-ja-filter
language:
- ja
- en
base_model:
- llm-jp/llm-jp-3-13b
---

# Model Card for Model ID

This is Full Parameter Fine Tuned model based on `llm-jp/llm-jp-3-13B`.
See the base details [here](https://huggingface.co/llm-jp/llm-jp-3-13b).

Made for the task of `elyza-tasks-100-TV` which Matsuo Lab made in a class.

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

- **Developed by:** [Yuto-24](https://github.com/Yuto-24/)
- **Model type:** Text Generation
- **Language(s) (NLP):** Japanese, English
- **License:** CC-BY-4.0
- **Finetuned from model:** [llm-jp/llm-jp-3-13B](https://huggingface.co/llm-jp/llm-jp-3-13b)

### Model Sources [optional]

<!-- Provide the basic links for the model. -->

- **Repository:** coming soon...

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

### Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

```txt:requirements.txt
numpy
torch>=2.3.0
datasets
transformers>=4.40.1
accelerate>=0.29.3
flash-attn>=2.5.8
FlagEmbedding
```

~~~python
import torch
import numpy as np

from datasets import Dataset, load_dataset
from FlagEmbedding import BGEM3FlagModel
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TextStreamer,
    BitsAndBytesConfig,
)

elyza_tasks_datasets = load_dataset("elyza/ELYZA-tasks-100")

model = BGEM3FlagModel("BAAI/bge-m3")
target_texts = elyza_tasks_datasets["test"]["input"].copy()
target_embeds = model.encode(target_texts)["dense_vecs"]


def retrieve(input_text):
    global target_embeds

    input_texts = [input_text]
    input_embeds = model.encode(input_texts)["dense_vecs"]

    # 類似度の計算
    similarity = input_embeds @ target_embeds.T
    most_similar_text = target_texts[np.argmax(similarity)]

    target_index = target_texts.index(most_similar_text)
    return target_index


class CallLLM:
    def __init__(self, model_name_or_path: str) -> None:
        self.model = AutoModelForCausalLM.from_pretrained(
            model_name_or_path,
            trust_remote_code=True,
            device_map="auto",
        ).eval()
        self.tokenizer = AutoTokenizer.from_pretrained(
            model_name_or_path,
            trust_remote_code=True,
        )
        self.streamer = TextStreamer(
            self.tokenizer,
        )
        self.call_type = None
        print(f"{self.model.device = }")

    def __call__(self, input_text: str, call_type: str = None, stream=False, **kwargs):
        self.call_type = call_type
        # print(f"Using call_type: {self.call_type}")

        call_type_dict = {
            "chat_template": self.__call_chat_template,
        }

        if self.call_type not in call_type_dict.keys():
            raise ValueError(
                f"Please set the call_type. You can select from {call_type_dict.keys()}"
            )
        output = call_type_dict[call_type](input_text.strip(), stream=stream, **kwargs)
        return output

    def merge_adapter(self, lora_adapter_path):
        # PEFTモデルとしてLoRAアダプターをベースモデルに結合
        self.model = PeftModel.from_pretrained(self.model, lora_adapter_path)
        self.model = self.model.merge_and_unload()

    def __call_chat_template(self, input_text: str = "", system_prompt: str = "あなたは、大塚商会の誠実で優秀なアシスタントです。", ** kwargs):
        prompt = []
        if system_prompt and system_prompt != "":
            prompt.append({"role": "system", "content": system_prompt})
        if input_text and input_text != "":
            prompt.append({"role": "user", "content": input_text})

        tokenized_input = self.tokenizer.apply_chat_template(
            prompt,
            return_tensors="pt",
        )

        output = self.__inference(tokenized_input, **kwargs)
        return output

        output = self.__inference(tokenized_input, **kwargs)
        return output

    def __inference(self, tokenized_input, stream: bool, **kwargs):
        tokenized_input = tokenized_input.to(self.model.device)
        attention_mask = torch.ones_like(tokenized_input)

        default_inference_params = {
            "attention_mask": attention_mask,
            "max_new_tokens": 512,
            "do_sample": False,
            "repetition_penalty": 1.2,
            "eos_token_id": self.tokenizer.eos_token_id,
            "pad_token_id": self.tokenizer.eos_token_id,
            # "eos_token_id": self.tokenizer.encode("<|im_end|>"),
        }

        inference_params = default_inference_params.copy()
        inference_params.update(**kwargs)
        if stream:
            inference_params.update(streamer=self.streamer)

        # Inference
        with torch.no_grad():
            outputs = self.model.generate(
                tokenized_input,
                **inference_params,
            )[0]
        output = self.tokenizer.decode(
            outputs[tokenized_input.size(1):],
            skip_special_tokens=True,
        )
        return output

model_path_or_id = "Yuto-24/llm-jp-3-13B-Tengentoppa_magpie"

# Loading model here.
llm = CallLLM(model_path_or_id)

SYSTEM_PROMPT = """
# あなたが必ず従うべき事項

## 役割

あなたは誠実で優秀なアシスタントです。
質問に対し、簡潔に答えます。
ハルシネーションをしません。
必ず正しい情報のみを答えます。

## 指示

- 評価観点に沿った出力を作成します。
- ユーザから特別な指示が与えられている場合には、必ず従います。
- 具体例には評価観点が含まれていますが、あなたが考える「出力」のみを回答してください。
- 評価観点は、人間があなたの出力を評価するために利用します。
- 論理的にステップバイステップで考えてください。

## 具体例

```markdown
{examples}
```
""".strip()

EXAMPLE_TEMPLATE = """
### 入力

{dataset_input}

### 評価観点

{dataset_eval_aspect}

### 出力

{dataset_answer}
""".strip()


# タスクとなるデータの読み込み
# omnicampusの開発環境では、左にタスクのjsonlをドラッグアンドドロップしてから実行

import os
import json

datasets = []
with open(f"{os.path.dirname(os.path.abspath('**file**'))}/workspace/elyza-tasks-100-TV_0.jsonl", "r") as f:
    item = ""
    for line in f:
        line = line.strip()
        item += line
        if item.endswith("}"):
            datasets.append(json.loads(item))
            item = ""

# モデルによるタスクの推論。
import re
from tqdm import tqdm

results = []
n = 2


for data in tqdm(datasets, smoothing=0.0):
    input_text = data["input"]
    dataset_index_list = retrieve(input_text, n)

    examples = ""
    for dataset_index in dataset_index_list:
        examples += EXAMPLE_TEMPLATE.format(
            dataset_input=elyza_tasks_datasets["test"]["input"][dataset_index].strip(),
            dataset_eval_aspect=elyza_tasks_datasets["test"]["eval_aspect"][dataset_index].strip(),
            dataset_answer=elyza_tasks_datasets["test"]["output"][dataset_index].strip(),
        )

    system_prompt = SYSTEM_PROMPT.format(
        examples=examples.strip(),
    )
    # print(examples)
    # print(input_text)

    output = llm(input_text=input_text,
                 system_prompt=system_prompt,
                 call_type="chat_template",
                 repetition_penalty=1.15,
                 # stream=True,
                 ).strip()
    # print("-----------------------------------------------------------------------------------------------------------------------------------")
    print(output.strip())
    print("===================================================================================================================================")
    print(re.sub(r"^[\s\S]*?### 出力", "", re.sub(r"^[\s\S]*?\*\*出力\*\*:", "", output)).strip())
    print("-----------------------------------------------------------------------------------------------------------------------------------")

    results.append({
        "task_id": data["task_id"],
        "input": input_text,
        "output_org": output.strip(),
        "output": re.sub(r"^[\s\S]*?### 出力", "", output).strip(),
        "elyza_tasks_id": dataset_index,
        "dataset_input": elyza_tasks_datasets["test"]["input"][dataset_index],
        "dataset_eval_aspect": elyza_tasks_datasets["test"]["eval_aspect"][dataset_index],
        "dataset_answer": elyza_tasks_datasets["test"]["output"][dataset_index],
    })

# results にタスクの解答が入っている

from pprint import pprint
import pandas as pd


# 最大表示「列」数の指定
pd.set_option("display.max_columns", 0)
# 最大表示「行」数の指定
pd.set_option("display.max_rows", 100)
pd.set_option("display.max_colwidth", 550)


json4df = {
    "task_id": [],
    "input": [],
    "output": [],
    "output_org": [],
    # "elyza_tasks_id": [],
    # "dataset_input": [],
    # "dataset_eval_aspect": [],
    # "dataset_answer": [],
}

for result in results:
    json4df["task_id"].append(result["task_id"])
    json4df["input"].append(result["input"])
    json4df["output_org"].append(result["output_org"])
    json4df["output"].append(result["output"])

JSON_FILE_NAME = "llm-jp-3-13B-Tengentoppa-FPFT-magpie-FPFT-elyza-RAG_v2"

result4out = results.copy()
results


# 本コードではinputとeval_aspectも含んでいますが、なくても問題ありません。
# 必須なのはtask_idとoutputとなります。

import re
import sys
from os.path import dirname, abspath, join, isfile


result4out = results.copy()


WD = dirname(abspath("__file__"))
json_dir = join(
    WD,
    "..",
    "jsonl",
)


if JSON_FILE_NAME != "":
    file_path = join(json_dir, f"{JSON_FILE_NAME}.jsonl")
else:
    jsonl_id = re.sub(".*/", "", merged_model_path)
    file_path = join(json_dir, f"{jsonl_id}-outputs.jsonl")

assert not isfile(file_path), f"Error: File `{file_path}` is already exist."

with open(file_path, "w", encoding="utf-8") as f:
    for result in result4out:
        result = {k: v for k, v in result.items() if k != "elyza_tasks_id" and k != "dataset_input" and k !=
                  "dataset_eval_aspect" and k != "dataset_answer"}
        json.dump(
            result, f, ensure_ascii=False
        )  # ensure_ascii=False for handling non-ASCII characters
        f.write("\n")


~~~

### Downstream Use [optional]

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

[More Information Needed]

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

[More Information Needed]

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

[More Information Needed]

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

## How to Get Started with the Model

Use the code below to get started with the model.

[More Information Needed]

## Training Details

### Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

- [DeL-TaiseiOzaki/Tengentoppa-sft-v1.0](https://huggingface.co/datasets/DeL-TaiseiOzaki/Tengentoppa-sft-v1.0)
- [llm-jp/magpie-sft-v1.0](https://huggingface.co/datasets/llm-jp/magpie-sft-v1.0)
- [ntotsuka123/clean3-ultraboros-20k-ja-filter](https://huggingface.co/datasets/ntotsuka123/clean3-ultraboros-20k-ja-filter)

### Training Procedure

using axolotl and yaml below.

```yaml: For the first training
base_model: llm-jp/llm-jp-3-13b
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

# domain_yyyymmdd
output_dir: outputs/matsuo/llm-jp/3/13B/FPFT_20241213

chat_template: chatml
default_system_message: あなたは、大塚商会の誠実で優秀なアシスタントです。

shuffle_merged_datasets: true
datasets:
  # # General
  # - path: data/general/magpie-sft-v1.0.jsonl
  #   ds_type: json
  #   type: chat_template
  #   chat_template: chatml
  #   field_messages: conversations
  #   message_field_role: role
  #   message_field_content: content
  #   roles:
  #     user:
  #       - user
  #     assistant:
  #       - assistant
  #     system:
  #       - system
  - path: data/general/Tengentoppa-sft-v1.0.jsonl
    ds_type: json
    type: alpaca
  # - path: data/general/clean3-ultraboros-20k-ja-filter_train.jsonl
  #   ds_type: json
  #   type: chat_template
  #   # chat_template: chatml
  #   field_messages: conversations
  #   message_field_role: role
  #   message_field_content: value
  #   roles:
  #     user:
  #       - human
  #     assistant:
  #       - gpt
  #     system:
  #       - system
  #   train_on_eos: turn

val_set_size: 0.05

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

# warmup_steps: 100
warmup_ratio: 0.1
evals_per_epoch: 1
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  eos_token: <|im_end|>
```

```yaml: For the second training
base_model: outputs/matsuo/llm-jp/3/13B/FPFT_20241213
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

# domain_yyyymmdd
output_dir: outputs/matsuo/llm-jp/3/13B/FPFT_20241215

chat_template: chatml
default_system_message: あなたは、大塚商会の誠実で優秀なアシスタントです。

shuffle_merged_datasets: true
datasets:
  - path: data/general/magpie-sft-v1.0.jsonl
    ds_type: json
    type: chat_template
    chat_template: chatml
    field_messages: conversations
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant
      system:
        - system
  # - path: data/general/Tengentoppa-sft-v1.0.jsonl
  #   ds_type: json
  #   type: alpaca
  - path: data/general/clean3-ultraboros-20k-ja-filter_train.jsonl
    ds_type: json
    type: chat_template
    chat_template: chatml
    field_messages: conversations
    message_field_role: role
    message_field_content: value
    roles:
      user:
        - human
      assistant:
        - gpt
      system:
        - system
    ## NOTE: Leaving the below empty will default to using the simple legacy tokenization strategy where only last message is trained on.
    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
    roles_to_train: ["gpt", "assistant"]
    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:
    # - all: train on all EOS tokens
    # - turn: train on the EOS token at the end of each trainable turn
    # - last: train on the last EOS token in the conversation
    train_on_eos: last

val_set_size: 0.05

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.00002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

# warmup_steps: 100
warmup_ratio: 0.1
evals_per_epoch: 1
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  eos_token: <|im_end|>
```

## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

### Testing Data, Factors & Metrics

#### Testing Data

<!-- This should link to a Dataset Card if possible. -->

[More Information Needed]

#### Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

[More Information Needed]

#### Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

[More Information Needed]

### Results

[More Information Needed]

#### Summary

## Model Examination [optional]

<!-- Relevant interpretability work for the model goes here -->

[More Information Needed]

## Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).

- **Hardware Type:** [More Information Needed]
- **Hours used:** [More Information Needed]
- **Cloud Provider:** [More Information Needed]
- **Compute Region:** [More Information Needed]
- **Carbon Emitted:** [More Information Needed]

## Technical Specifications [optional]

### Model Architecture and Objective

[More Information Needed]

### Compute Infrastructure

[More Information Needed]

#### Hardware

[More Information Needed]

#### Software

[More Information Needed]

## Citation [optional]

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

[More Information Needed]

**APA:**

[More Information Needed]

## Glossary [optional]

<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->

[More Information Needed]

## More Information [optional]

[More Information Needed]

## Model Card Authors [optional]

[More Information Needed]

## Model Card Contact

[More Information Needed]