We introduce AceInstruct, a family of advanced SFT models for coding, mathematics, and general-purpose tasks. The AceInstruct family, which includes AceInstruct-1.5B, 7B, and 72B, is Improved using Qwen. These models are fine-tuned on Qwen2.5-Base using general SFT datasets. These same datasets are also used in the training of AceMath-Instruct. Different from AceMath-Instruct which is specialized for math questions, AceInstruct is versatile and can be applied to a wide range of domains. Benchmark evaluations across coding, mathematics, and general knowledge tasks demonstrate that AceInstruct delivers performance comparable to Qwen2.5-Instruct.

For more information about AceInstruct, check our website and paper.

Benchmark Results

	Qwen2.5-1.5B-Instruct	AceInstruct-1.5B	Qwen2.5-7B-Instruct	AceInstruct-7B	Qwen2.5-72B-Instruct	AceInstruct-72B
HumanEval	61.60	73.17	84.80	85.37	86.60	89.63
MBPP	63.20	65.76	79.20	74.32	88.20	83.66
GSM8K	73.20	80.44	91.60	93.10	95.80	96.36
MATH	55.20	60.34	75.50	76.40	83.10	84.50
MMLU	58.37	58.17	74.51	74.68	84.67	83.88
MMLU Pro	32.40	33.78	56.30	54.50	71.10	66.10
Average	57.33	61.94	76.99	76.40	84.91	84.02

We compare AceInstruct to Qwen2.5-Instruct across coding, mathematics, and general knowledge tasks. We find that AceInstruct-1.5B outperforms Qwen2.5-1.5B-Instruct (61.94 vs. 57.33), while AceInstruct-7B and AceInstruct-72B perform similarly to Qwen2.5-7B-Instruct and Qwen2.5-72B-Instruct.

All Resources

How to use

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "AceInstruct-72B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")

prompt = "Tell me something about artificial intelligence."
messages = [{"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to("cuda")

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=1024
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Correspondence to

Zihan Liu (zihanl@nvidia.com), Yang Chen (yachen@nvidia.com), Wei Ping (wping@nvidia.com)

Citation

If you find our work helpful, we’d appreciate it if you could cite us.

@article{acemath2024, title={AceMath: Advancing Frontier Math Reasoning with Post-Training and Reward Modeling}, author={Liu, Zihan and Chen, Yang and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei}, journal={arXiv preprint}, year={2024} }

License

All models in the AceInstruct family are for non-commercial use only, subject to Terms of Use of the data generated by OpenAI. We put the AceInstruct models under the license of Creative Commons Attribution: Non-Commercial 4.0 International.

nvidia/AceInstruct-72B

Introduction

Benchmark Results

All Resources

AceMath Instruction Models

AceMath Reward Models

Evaluation & Training Data

General Instruction Models

How to use

Correspondence to

Citation

License

nvidia/
AceInstruct-72B