digital-socrates-7b / README.md
SaveBertAndGpt's picture
Update citation
3150038 verified
metadata
language: en
license: apache-2.0
library_name: transformers
model-index:
  - name: digital-socrates-7b
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 54.44
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=allenai/digital-socrates-7b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 75.99
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=allenai/digital-socrates-7b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 51.41
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=allenai/digital-socrates-7b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 44.88
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=allenai/digital-socrates-7b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 73.09
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=allenai/digital-socrates-7b
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 17.89
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=allenai/digital-socrates-7b
          name: Open LLM Leaderboard

This is the Digital Socrates 7B (DS-7B) model described in our paper: Digital Socrates: Evaluating LLMs through explanation critiques (ACL Anthology link: https://aclanthology.org/2024.acl-long.302, arXiv link: https://arxiv.org/abs/2311.09613).

The recommended, better performing 13B model can be found at https://huggingface.co/allenai/digital-socrates-13b

The DS-7B model is a fine-tuned version of Llama-2-7b-Chat, please review its guidelines and licensing before use.

What is Digital Socrates?

Digital Socrates is an open-source, automatic explanation-critiquing model. It is useful for revealing insights about student models by examining their reasoning chains. The use of Digital Socrates allows for nuanced, interpretable automatic evaluation of explanations without expensive API calls or human annotations.

While LLMs can provide reasoned explanations along with their answers, the nature and quality of those explanations are still poorly understood. In response, our goal is to define a detailed way of characterizing the explanation capabilities of modern models and to create a nuanced, interpretable explanation evaluation tool that can generate such characterizations automatically.

Given a question (together with the multiple-choice options and gold answer),

When Dennis operates his lawnmower, he notices the engine makes a lot of noise. He also notices that the engine gets very hot. Which best describes the heat and noise generated from the lawnmower? 
(A) a change in phase 
(B) thermal expansion 
(C) an increase in entropy 
(D) mechanical advantage 

Answer sheet: (C)

as well as a student model explanation and answer,

1) The question states that the lawnmower engine makes a lot of noise.
2) The question states that the lawnmower engine gets very hot.
3) Noise and heat are both forms of energy.
4) The noise and heat generated from the lawnmower are a result of the conversion of energy from the fuel to mechanical energy.

Answer: (D)

Digital Socrates gives a critique of the model-generated explanation that provides localized feedback on the most significant flaw (if any) in the explanation, suggestions to address the identified flaw, as well as a numeric rating indicating the explanation quality:

The explanation states or suggests the following:
* Main flaw (standalone statement): "The noise and heat generated from the lawnmower are a result of the conversion of energy from the fuel to mechanical energy."
* Dimension: incorrect_information

Consider these points for revising the explanation:
* General: Remember that noise and heat are not forms of energy. They are byproducts of energy conversion.
* Specific: In this case, the noise and heat generated by the lawnmower are not a result of the conversion of energy from the fuel to mechanical energy. They are byproducts of the combustion process.

Explanation score: 2

Remarkably, despite being orders of magnitude smaller than GPT-4, our Digital Socrates models are capable of generating critiques close to GPT-4 critiques in terms of human rating and other quantitative measures (correlation of explanation scores given and error category matches). Through quantitative and qualitative analysis, we demonstrate how Digital Socrates is useful for revealing insights about student models by examining their reasoning chains.

We invite you to try out Digital Socrates for your own application!

How to use Digital Socrates?

We provide a quick example of how you can try out Digital Socrates with just a few lines of code:

'DSCritiqueBank-V1' used below can be downloaded from our dataset page.

import json
from transformers import AutoTokenizer, AutoModelForCausalLM
# Load model and tokenizer
model_path = "allenai/digital-socrates-7b"
model = AutoModelForCausalLM.from_pretrained(model_path).to("cuda:0")
tokenizer = AutoTokenizer.from_pretrained(model_path)

# Define input data
question = "When Dennis operates his lawnmower, he notices the engine makes a lot of noise. He also notices that the engine gets very hot. Which best describes the heat and noise generated from the lawnmower? (A) a change in phase (B) thermal expansion (C) an increase in entropy (D) mechanical advantage"
explanation = "1) The question states that the lawnmower engine makes a lot of noise.\n2) The question states that the lawnmower engine gets very hot.\n3) Noise and heat are both forms of energy.\n4) The noise and heat generated from the lawnmower are a result of the conversion of energy from the fuel to mechanical energy."
answerkey = "C"
predictedanswer = "D"

# construct prompt (Llama conventions)
with open("../DSCritiqueBank-V1/DSCB-prompts.json") as file:
    prompts = json.load(file)

system_prompt = prompts['digital_socrates_v1']['system']
user_prompt = prompts['digital_socrates_v1']['main'].replace("[[QUESTION]]", question).replace("[[EXPLANATION]]", explanation).replace("[[PREDICTEDANSWER]]", predictedanswer).replace("[[ANSWERKEY]]", answerkey)

full_prompt = f"[INST] <<SYS>>\n{system_prompt}\n<</SYS>{user_prompt} [/INST]\n\n"

# Run model
input_ids = tokenizer.encode(full_prompt, return_tensors="pt").to("cuda:0")
output = model.generate(input_ids, max_new_tokens=512, temperature=0)
res = tokenizer.batch_decode(output, skip_special_tokens=True)

Print the output:

>>> print(res[0].split("[/INST]")[-1])

The explanation states or suggests the following:
* Main flaw (standalone statement): "The noise and heat generated from the lawnmower are a result of the conversion of energy from the fuel to mechanical energy."
* Dimension: incorrect_information

Consider these points for revising the explanation:
* General: Remember that noise and heat are not forms of energy. They are byproducts of energy conversion.
* Specific: In this case, the noise and heat generated by the lawnmower are not a result of the conversion of energy from the fuel to mechanical energy. They are byproducts of the combustion process.

Explanation score: 2

More details about Digital Socrates ...

For more details about Digital Socrates, please refer to our:

Citation

@inproceedings{gu-etal-2024-digital,
    title = "Digital Socrates: Evaluating {LLM}s through Explanation Critiques",
    author = "Gu, Yuling  and
      Tafjord, Oyvind  and
      Clark, Peter",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.302",
    pages = "5559--5586",
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 52.95
AI2 Reasoning Challenge (25-Shot) 54.44
HellaSwag (10-Shot) 75.99
MMLU (5-Shot) 51.41
TruthfulQA (0-shot) 44.88
Winogrande (5-shot) 73.09
GSM8k (5-shot) 17.89