MeGan (meta_swiglu) - CrossFit & UnifiedQA SFT
This model is a fine-tuned version of Llama-3.1-8B-Instruct implementing the meta-gating mechanism proposed in Learn-To-Learn on Arbitrary Textual Conditioning: A Hypernetwork-Driven Meta-Gated LLM.
- GitHub Repository: MeGan
- Dataset: CrossFit and UnifiedQA (Hugging Face Dataset)
Model description
Conventional LLMs may suffer from corpus heterogeneity and subtle condition changes. This work activates the meta-signal of $\beta$ within the SwiGLU blocks, resulting in a meta-gating mechanism that adaptively adjusts the nonlinearity of FFN. A hypernetwork is employed which dynamically produces $\beta$ on textual conditions, providing meta-controllability on LLMs.
This checkpoint is trained on a subset (non_nli_to_nli) of CrossFit and UnifiedQA.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 32
- total_train_batch_size: 32
- total_eval_batch_size: 256
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 1.0
Framework versions
- Transformers 4.46.1
- Pytorch 2.3.0+cu121
- Datasets 2.20.0
- Tokenizers 0.20.3
Citation
If you find this work useful, please consider citing:
@article{ji2026learntolearn,
title={Learn-to-learn on Arbitrary Textual Conditioning: A Hypernetwork-Driven Meta-Gated LLM},
author={Luo Ji and Qi Qin and Ningyuan Xi and Teng Chen and Qingqing Gu and Hongyan Li},
journal={arXiv preprint arXiv:2605.01973},
year={2026}
}