Exllamav2 quant (exl2 / 4.25 bpw) made with ExLlamaV2 v0.0.21

Other EXL2 quants:

Quant	Model Size	lm_head
2.2	12762 MB	6
2.5	14191 MB	6
3.0	16931 MB	6
3.5	19724 MB	6
3.75	21098 MB	6
4.0	22488 MB	6
4.25	23875 MB	6
5.0	28028 MB	6
6.0	33588 MB	8
6.5	36031 MB	8
8.0	41342 MB	8

🦣 MAmmoTH2: Scaling Instructions from the Web

Project Page: https://tiger-ai-lab.github.io/MAmmoTH2/

Paper: https://arxiv.org/pdf/2405.03548

Code: https://github.com/TIGER-AI-Lab/MAmmoTH2

Introduction

Introducing 🦣 MAmmoTH2, a game-changer in improving the reasoning abilities of large language models (LLMs) through innovative instruction tuning. By efficiently harvesting 10 million instruction-response pairs from the pre-training web corpus, we've developed MAmmoTH2 models that significantly boost performance on reasoning benchmarks. For instance, MAmmoTH2-7B (Mistral) sees its performance soar from 11% to 34% on MATH and from 36% to 67% on GSM8K, all without training on any domain-specific data. Further training on public instruction tuning datasets yields MAmmoTH2-Plus, setting new standards in reasoning and chatbot benchmarks. Our work presents a cost-effective approach to acquiring large-scale, high-quality instruction data, offering a fresh perspective on enhancing LLM reasoning abilities.

	Base Model	MAmmoTH2	MAmmoTH2-Plus
7B	Mistral	🦣 MAmmoTH2-7B	🦣 MAmmoTH2-7B-Plus
8B	Llama-3	🦣 MAmmoTH2-8B	🦣 MAmmoTH2-8B-Plus
8x7B	Mixtral	🦣 MAmmoTH2-8x7B	🦣 MAmmoTH2-8x7B-Plus

Training Data

Please refer to https://huggingface.co/datasets/TIGER-Lab/WebInstructSub for more details.

Training Procedure

The models are fine-tuned with the WEBINSTRUCT dataset using the original Llama-3, Mistral and Mistal models as base models. The training procedure varies for different models based on their sizes. Check out our paper for more details.

Evaluation

The models are evaluated using open-ended and multiple-choice math problems from several datasets. Here are the results:

Model	TheoremQA	MATH	GSM8K	GPQA	MMLU-ST	BBH	ARC-C	Avg
MAmmoTH2-7B (Updated)	29.0	36.7	68.4	32.4	62.4	58.6	81.7	52.7
MAmmoTH2-8B (Updated)	30.3	35.8	70.4	35.2	64.2	62.1	82.2	54.3
MAmmoTH2-8x7B	32.2	39.0	75.4	36.8	67.4	71.1	87.5	58.9
MAmmoTH2-7B-Plus (Updated)	31.2	46.0	84.6	33.8	63.8	63.3	84.4	58.1
MAmmoTH2-8B-Plus (Updated)	31.5	43.0	85.2	35.8	66.7	69.7	84.3	59.4
MAmmoTH2-8x7B-Plus	34.1	47.0	86.4	37.8	72.4	74.1	88.4	62.9

To reproduce our results, please refer to https://github.com/TIGER-AI-Lab/MAmmoTH2/tree/main/math_eval.

Chat Format

The template used to build a prompt for the Instruct model is defined as follows:

<s> [INST] Instruction [/INST] Model answer</s> [INST] Follow-up instruction [/INST]

Note that <s> and </s> are special tokens for beginning of string (BOS) and end of string (EOS) while [INST] and [/INST] are regular strings.

But we also found that the model is not very sensitive to the chat template.

Usage

You can use the models through Huggingface's Transformers library. Use the pipeline function to create a text-generation pipeline with the model of your choice, then feed in a math problem to get the solution. Check our Github repo for more advanced use: https://github.com/TIGER-AI-Lab/MAmmoTH2

Limitations

We've tried our best to build math generalist models. However, we acknowledge that the models' performance may vary based on the complexity and specifics of the math problem. Still not all mathematical fields can be covered comprehensively.

Citation

If you use the models, data, or code from this project, please cite the original paper:

@article{yue2024mammoth2,
  title={MAmmoTH2: Scaling Instructions from the Web},
  author={Yue, Xiang and Zheng, Tuney and Zhang, Ge and Chen, Wenhu},
  journal={arXiv preprint arXiv:2405.03548},
  year={2024}
}