notHumpback-M0

This model follows the Humpback architecture, proposed in the paper Self-Alignment with Instruction Backtranslation by Li et al.

It represents the "seed model", which is trained on a small amount of gold data and then used to score the instruction-response pairs generated by the "backward model".

Humpback uses instruction backtranslation on a web corpus to generate input-output pairs (self-augmentation), creating a richer dataset for fine-tuning models without the need for additional manual annotation. The model then iteratively curates the created dataset, scoring the pairs by quality, and is then finetuned on the resulting subset of all pairs with the highest possible score (self-curation).

Varying from the original paper, this model is a fine-tuned version of meta-llama/Llama-3.2-3B. It has been trained using TRL.

The dataset used to train this model has been sampled from the oasst1 dataset. To enable the model to judge and score the generated pairs, the model undergoes basic instruction-tuning on the input-output pairs contained in the dataset.

Framework versions

• TRL: 0.12.1
• Transformers: 4.46.3
• Pytorch: 2.5.1
• Datasets: 3.1.0
• Tokenizers: 0.20.3

Citations

Original paper:

@misc{li2023selfalignment,
    title={Self-Alignment with Instruction Backtranslation},
    author={Xian Li and Ping Yu and Chunting Zhou and Timo Schick and Luke Zettlemoyer and Omer Levy and Jason Weston and Mike Lewis},
    year={2023},
    eprint={2308.06259},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallouédec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}