Skywork-Critic-Llama-3.1-70B

<div align="center"> <img src="misc/misc_fig.jpg" width="400"/> 🤗 <a href="https://huggingface.co/Skywork" target="_blank">Hugging Face</a> • 🤖 <a href="https://modelscope.cn/organization/Skywork" target="_blank">ModelScope</a> <br> <br> <br> </div>

Introduction to Skywork Critic Series Models

Skywork-Critic-Llama3.1-70B and Skywork-Critic-Llama3.1-8B, developed by the SkyworkAI Alignment Team, are advanced judge models that excel at pairwise preference evaluation. These models compare and assess input pairs, offering nuanced judgments on their relative quality or suitability. Leveraging their deep understanding of language and context, Skywork-Critic models provide valuable insights for various applications, including data improvement, evaluation, and reward modeling.

Training Details

Skywork-Critic-Llama3.1-70B and Skywork-Critic-Llama3.1-8B are built on Meta Llama-3.1-70B-Instruct and Llama-3.1-8B-Instruct respectively. These models have undergone fine-tuning using a diverse array of high-quality datasets, including:

• Cleaned open-source data: We utilize a high-quality subset of HelpSteer2, OffsetBias, WildGuard (adversarial) and Magpie DPO series(Ultra,Pro (Llama-3.1),Pro,Air). For more details, please refer to our Skywork-Reward-Preference-80K-v0.1 dataset. Additionally, we integrate several open-source, high-quality critic datasets such as Open-Critic-GPT into our training process.
• In-house human annotation data: This includes both pointwise scoring across many dimensions for a single response and pairwise comparisons between two responses. Each dimension incorporates a rationale for the assigned score. Please note that manually labeled data is very expensive to obtain. We only have a few hundred manually labeled data points, all of which are in Chinese, so the ability to perform single rating might not be particularly strong.
• Synthetic critic data: We use a similar appoarch to self-taught. Specifically, we employed two methods to generate inferior responses for a given instruction: 1) Creating a similar instruction and then generating a response for this new instruction. 2) Introducing subtle errors into high-quality responses.
• Critic-related chat data: We incorporate critic-related chat data to maintain the model's conversational capabilities.

The training employs instruction-tuning methodology, focusing on pairwise preference evaluation and general chat tasks. We have conducted a thorough verification process to ensure our training dataset does not contain any test set information from RewardBench, maintaining the integrity of our evaluation results.

RewardBench Leaderboard for Generative Models

We evaluate our models on RewardBench using the official test script.

As of September 2024, Skywork-Critic-Llama3.1-70B ranks first on RewardBench for generative models across all sizes, while Skywork-Critic-Llama3.1-8B tops the list for generative models under 10B parameters. (Note: An asterisk (*) indicates an open-source model.)

Demo Code

Below are two examples of how to use the Skywork Critic model: as a preference data selector, and as a judge to generate scores and rationales for instruction-response pairs.

Skywork Critic Model as a Preference Data Selector

Here is an example showing how to use the Skywork Critic Model as a preference data selector. It distinguishes between chosen and rejected training data for Direct Preference Optimization (DPO) training.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# An Example Case
prompt = "Jane has 12 apples. She gives 4 apples to her friend Mark, then buys 1 more apple, and finally splits all her apples equally among herself and her 2 siblings. How many apples does each person get?"
responseA = "1. Jane starts with 12 apples and gives 4 to Mark. 12 - 4 = 8. Jane now has 8 apples.\n2. Jane buys 1 more apple. 8 + 1 = 9. Jane now has 9 apples.\n3. Jane splits the 9 apples equally among herself and her 2 siblings (3 people in total). 9 ÷ 3 = 3 apples each. Each person gets 3 apples."
responseB = "1. Jane starts with 12 apples and gives 4 to Mark. 12 - 4 = 8. Jane now has 8 apples.\n2. Jane buys 1 more apple. 8 + 1 = 9. Jane now has 9 apples.\n3. Jane splits the 9 apples equally among her 2 siblings (2 people in total). 9 ÷ 2 = 4.5 apples each. Each person gets 4 apples."

# feed a natural language prompt to generative model
prompt_template = """Please act as an impartial judge and evaluate the quality of the responses provided by two AI assistants to the user question displayed below. You should choose the assistant that follows the user\'s instructions and answers the user\'s question better. 
Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of their responses. Avoid any position biases and ensure that the order in which the responses were presented does not influence your decision. Do not allow the length of the responses to influence your evaluation. Do not favor certain names of the assistants. Be as objective as possible. 
Please directly output your final verdict by strictly following this format: "[[A]]" if assistant A is better, "[[B]]" if assistant B is better.

[User Question]
{input}

[The Start of Assistant A's Answer]
{response_a}
[The End of Assistant A's Answer]

[The Start of Assistant B's Answer]
{response_b}
[The End of Assistant B's Answer]
"""

user_message = prompt_template.format(input=prompt, response_a=responseA, response_b=responseB)

conversation = [{"role": "user", "content": user_message}]

model_name = "Skywork/Skywork-Critic-Llama3.1-70B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

input_ids = tokenizer.apply_chat_template(
    conversation, 
    tokenize=True, 
    add_generation_prompt=True,
    return_tensors="pt").to(model.device)

generation = model.generate(
    input_ids=input_ids,
    max_new_tokens=2048,
    do_sample=False,
    pad_token_id=128009,
    temperature=0)

completion = tokenizer.decode(
    generation[0][len(input_ids[0]):], 
    skip_special_tokens=True, 
    clean_up_tokenization_spaces=True)

print(completion)

# Output:
# The generative model should output "[[A]]"

Skywork Critic Model as a Judge

Here is an example showing how to use the Skywork Critic model as a judge. For an instruction-response pair, the Skywork-Critic Model generates a score and rationale based on specific evaluation criteria. Our preliminary research indicates that 8B-parameter models struggle to produce reliable judgments for responses. Consequently, we exclusively utilize the Skywork-Critic-Llama3.1-70B model as our judge.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# An Example Case
prompt = "Jane has 12 apples. She gives 4 apples to her friend Mark, then buys 1 more apple, and finally splits all her apples equally among herself and her 2 siblings. How many apples does each person get?"

# Chosen Response
responseA = "1. Jane starts with 12 apples and gives 4 to Mark. 12 - 4 = 8. Jane now has 8 apples.\n2. Jane buys 1 more apple. 8 + 1 = 9. Jane now has 9 apples.\n3. Jane splits the 9 apples equally among herself and her 2 siblings (3 people in total). 9 ÷ 3 = 3 apples each. Each person gets 3 apples."

# Rejected Response
responseB = "1. Jane starts with 12 apples and gives 4 to Mark. 12 - 4 = 8. Jane now has 8 apples.\n2. Jane buys 1 more apple. 8 + 1 = 9. Jane now has 9 apples.\n3. Jane splits the 9 apples equally among her 2 siblings (2 people in total). 9 ÷ 2 = 4.5 apples each. Each person gets 4 apples."

# feed a natural language prompt to generative model
single_rating_prompt_template = """请扮演一个专业的对话质量评价专家，能够从多个维度诊断和分析 AI 助手对用户问题的回答，并进行总体打分（分值范围是 1-5）。你的评估应考虑回答的有用性、相关性、准确性、深度、创造性、安全性等众多维度，请注意，不同任务类型的指令对评估分析维度的侧重不一样，需要根据具体的问题进行分析。

[用户问题]
{input}

[助手回答的开始]
{response_a}
[助手回答的结束]

你的详细评估和总体打分为: """

# Critic responseA and responseB separately.
user_message = single_rating_prompt_template.format(input=prompt, response_a=responseA)
# user_message = single_rating_prompt_template.format(input=prompt, response_a=responseB)


conversation = [{"role": "user", "content": user_message}]

print(f"conversation => {conversation}")


model_name = "70B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

input_ids = tokenizer.apply_chat_template(
    conversation, 
    tokenize=True, 
    add_generation_prompt=True, 
    return_tensors="pt").to(model.device)

generation = model.generate(
    input_ids=input_ids,
    max_new_tokens=2048,
    do_sample=False,
    pad_token_id=128009,
    temperature=0)

completion = tokenizer.decode(
    generation[0][len(input_ids[0]):], 
    skip_special_tokens=True, 
    clean_up_tokenization_spaces=True)


print(completion)

Model-Generated Judgment for Chosen Response A

**评估分析：**

1. **有用性**：回答直接解决了用户的问题，提供了清晰的步骤和计算过程，非常有用。
2. **相关性**：回答完全相关于用户的问题，没有偏离主题。
3. **准确性**：计算过程准确无误，结果正确。
4. **深度**：回答提供了详细的步骤和解释，帮助用户理解问题的解决过程。
5. **创造性**：虽然回答是直接的计算过程，但在解释每一步时使用了简单的语言和明确的计算，使得回答易于理解。
6. **安全性**：回答中没有包含任何可能引起误解或危险的信息。

**总体打分：5**

**总结**：这个回答非常清晰、准确且有用，完全满足了用户的需求。它通过逐步解释和计算，帮助用户理解问题的解决过程。没有发现任何需要改进的地方，因此给予满分。

Model-Generated Judgment for Rejected Response B

**评估分析：**

1. **有用性**：回答直接解决了用户的问题，提供了每个人最终得到的苹果数量。
2. **相关性**：回答紧密相关于用户的问题，没有偏离主题。
3. **准确性**：回答中的计算过程是正确的，但在最后一步中出现了错误。Jane 分苹果时，应该是将苹果分给自己和她的2个兄弟姐妹，总共3个人，而不是2个人。
4. **深度**：回答提供了详细的计算步骤，帮助用户理解问题的解决过程。
5. **创造性**：回答是直接的，没有特别的创造性，但对于这个简单的数学问题来说，直接的回答是合适的。
6. **安全性**：回答没有包含任何可能引起误解或危险的信息。

**总体打分：** 4

**改进建议：**
- 在最后一步中，正确地计算每个人得到的苹果数量。Jane 应该将 9 个苹果分给自己和她的 2 个兄弟姐妹，总共 3 个人。因此，每个人得到的苹果数量应该是 9 ÷ 3 = 3 个苹果。

Declaration and License Agreement

Declaration

We hereby declare that the Skywork model should not be used for any activities that pose a threat to national or societal security or engage in unlawful actions. Additionally, we request users not to deploy the Skywork model for internet services without appropriate security reviews and records. We hope that all users will adhere to this principle to ensure that technological advancements occur in a regulated and lawful environment.

We have done our utmost to ensure the compliance of the data used during the model's training process. However, despite our extensive efforts, due to the complexity of the model and data, there may still be unpredictable risks and issues. Therefore, if any problems arise as a result of using the Skywork open-source model, including but not limited to data security issues, public opinion risks, or any risks and problems arising from the model being misled, abused, disseminated, or improperly utilized, we will not assume any responsibility.

License Agreement

The community usage of Skywork model requires Skywork Community License. The Skywork model supports commercial use. If you plan to use the Skywork model or its derivatives for commercial purposes, you must abide by terms and conditions within Skywork Community License.

Contact

If you have any questions or feedback, don't hesitate to reach out to our friendly team at [email protected] or [email protected]. Liang Zhao leads this project.

Citation

If you find our work helpful, please feel free to cite us using the following BibTeX entry:

@misc{skyworkcritic2024,
  title={Skywork Critic Model Series},
  author={Shiwen, Tu and Liang, Zhao and Liu, Chris Yuhao and Zeng, Liang and Liu, Yang},
  year={2024},
  month={September},
  howpublished={\url{https://huggingface.co/Skywork}},
  url={https://huggingface.co/Skywork},
}