metadata

license: apache-2.0
datasets:
  - dyyyyyyyy/ScaleQuest-Math
language:
  - en
metrics:
  - accuracy
library_name: transformers
pipeline_tag: text-generation

Unleashing Reasoning Capability of LLMs via Scalable Question Synthesis from Scratch

Model Card for Llama3-8B-ScaleQuest

We introduce ScaleQuest, a scalable and novel data synthesis method that utilizes small-size open-source models to generate questions from scratch without the need for seed data with complex augmentation constraints.

📑 Project Page: https://scalequest.github.io
💻 Code: https://github.com/yyDing1/ScaleQuest
📖 Paper: Unleashing Reasoning Capability of LLMs via Scalable Question Synthesis from Scratch
💾 Models in the 🤗 HuggingFace Hub: ScaleQuest-Models

Datasets & Models

Math Dataset: link

We release two question generator models and four problem-solving models.

Model	Type	MATH	Olympiad Bench	🤗 HuggingFace Download Link
ScaleQuest-DeepSeekMath-7B-QGen	question generator	-	-	link
ScaleQuest-Qwen2-Math-7B-QGen	question generator	-	-	link
Mistral-7B-ScaleQuest	problem solver	62.9	26.8	link
Llama3-8B-ScaleQuest	problem solver	64.4	25.3	link
DeepSeekMath-7B-ScaleQuest	problem solver	66.6	29.9	link
Qwen2-Math-7B-ScaleQuest	problem solver	73.4	38.5	link

Demo usage

Below is an example using Llama3-8B-ScaleQuest

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "dyyyyyyyy/Llama3-8B-ScaleQuest"

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

question = "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$."

sys_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request." + "\n\n"
query_prompt = "### Instruction:" + "\n"
# {query}
prompt_after_query = "\n\n"
resp_prompt = "### Response:" + "\n"
prompt_before_resp = ""
# {resp}
delim = "\n\n"

prefix_prompt = f"{query_prompt}{question}{prompt_after_query}{resp_prompt}{prompt_before_resp}".rstrip(" ")
full_prompt = sys_prompt + delim.join([prefix_prompt])

# print(full_prompt)

inputs = tokenizer(full_prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True))

Citation

@article{ding2024unleashing,
    title={Unleashing Reasoning Capability of LLMs via Scalable Question Synthesis from Scratch}, 
    author={Ding, Yuyang and Shi, Xinyu and Liang, Xiaobo and Li, Juntao and Zhu, Qiaoming and Zhang, Min},
    journal={https://arxiv.org/abs/2410.18693}, 
    year={2024}
}