Superswallow-70b-v0.1

Known Performance Issues

Two potential bugs have been found in this model:

NEED repetition_penalty
NEED high temperature

Reference: Japanese LLM benchmark results at Nejumi LLM Leaderboad Neo

The current benchmark results are worse than Swallow, which was used as a merge-based model. These bugs may be the reason for the low score, but it is also quite possible that this merged model is underperforming.

The author's experience is that the performance has improved with the test parameters. In particular, the ability to follow instructions during role play should be about 25% to 50% of the performance of the Tulu 2 DPO. I would be happy if you could try it out and give me feedback.

These issue may be caused by the self-attention layers and will be fixed in the next version. (However, I don't know if the benchmark score will improve)

Important Notice:

This model partially utilizes the parameters of Tulu V2 DPO finetuned based on Llama 2, so it may inherit the AI2 ImpACT license. Please use the model keeping in mind that there may be changes regarding the license if AI2 contacts me.

The AI2 ImpACT license includes information about data artifacts and model artifacts, but does not cover the case of directly applying parts of the LLM parameters of a model artifact to other models. However, I respect their research and great work, so I will change the license immediately if AI2 contacts me.

Description

This is a merge of pre-trained language models created using mergekit. The model was created by injecting the ability to follow user intent from Tulu 2 DPO into the Swallow instract model.

It was a proof of concept for merging LLMs trained in other languages, and paid close attention to preserving the linguistic capabilities of the merge-based model.

As far as I know, Swallow is the full set Llama 2 model(7B, 13B, 70B) that can output the most beautiful Japanese. Therefore, I used it as the base model for merging this time. Thank you for their wonderful work.

Test environment

This model was tested using text-generation-webui. I use preset simple-1 for Generation.

Users reported that setting repetition_penalty is important to prevent repeated output. If you run into any issues, be sure to check your settings. Additionally, a bug was discovered that caused an error at low temperatures.

temperature: 0.7
top_p: 0.9
repetition_penalty: 1.15
top_k: 20

Prompt template: Swallow (Alpaca format)

以下に、あるタスクを説明する指示があり、それに付随する入力が更なる文脈を提供しています。リクエストを適切に完了するための回答を記述してください。

### 指示:
{instruction}

### 応答:

Use the instruct model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "nitky/Superswallow-70b-v0.1"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, device_map="auto", load_in_4bit = True)


PROMPT_DICT = {
    "prompt_input": (
        "以下に、あるタスクを説明する指示があり、それに付随する入力が更なる文脈を提供しています。"
        "リクエストを適切に完了するための回答を記述してください。\n\n"
        "### 指示:\n{instruction}\n\n### 入力:\n{input}\n\n### 応答:"

    ),
    "prompt_no_input": (
        "以下に、あるタスクを説明する指示があります。"
        "リクエストを適切に完了するための回答を記述してください。\n\n"
        "### 指示:\n{instruction}\n\n### 応答:"
    ),
}

def create_prompt(instruction, input=None):
    """
    Generates a prompt based on the given instruction and an optional input.
    If input is provided, it uses the 'prompt_input' template from PROMPT_DICT.
    If no input is provided, it uses the 'prompt_no_input' template.

    Args:
        instruction (str): The instruction describing the task.
        input (str, optional): Additional input providing context for the task. Default is None.

    Returns:
        str: The generated prompt.
    """
    if input:
        # Use the 'prompt_input' template when additional input is provided
        return PROMPT_DICT["prompt_input"].format(instruction=instruction, input=input)
    else:
        # Use the 'prompt_no_input' template when no additional input is provided
        return PROMPT_DICT["prompt_no_input"].format(instruction=instruction)

# Example usage
instruction_example = "以下のトピックに関する詳細な情報を提供してください。"
input_example = "東京工業大学の主なキャンパスについて教えてください"
prompt = create_prompt(instruction_example, input_example)

input_ids = tokenizer.encode(
    prompt,
    add_special_tokens=False,
    return_tensors="pt"
)

tokens = model.generate(
    input_ids.to(device=model.device),
    max_new_tokens=200,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.15,
    top_k=20,
    do_sample=True,
)

out = tokenizer.decode(tokens[0], skip_special_tokens=True)
print(out)

Merge Details

Merge Method

This model was merged using the DARE TIES merge method using tokyotech-llm/Swallow-70b-instruct-hf as a base.

Models Merged

The following models were included in the merge:

allenai/tulu-2-dpo-70b

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: tokyotech-llm/Swallow-70b-instruct-hf 
    # no parameters necessary for base model
  - model: allenai/tulu-2-dpo-70b # follow user intent 
    parameters:
      density: 1
      weight:
      - filter: mlp.down_proj
        value: [0.3, 0.25, 0.25, 0.15, 0.1]
      - filter: mlp.gate_proj
        value: [0.7, 0.25, 0.5, 0.45, 0.4]
      - filter: mlp.up_proj
        value: [0.7, 0.25, 0.5, 0.45, 0.4]
      - filter: self_attn
        value: [0.7, 0.25, 0.5, 0.45, 0.4]
      - value: 0 # fallback for rest of tensors.
merge_method: dare_ties
base_model: tokyotech-llm/Swallow-70b-instruct-hf
dtype: bfloat16
tokenizer_source: union

nitky
/

Superswallow-70b-v0.1