Show HN: Benchmarks of continual learning model trained for a month on a laptop

A month ago, I shared mini-AGI, an experiment in training a byte-level, continually learning language model from scratch on a single 8 GB laptop GPU (an RTX 3070). To recap the setup briefly: it completely decouples model size from VRAM by paging experts to and from an NVMe drive (currently 397 experts / 1.26B total parameters on disk, with only 32 resident / ~109M active in 5.1 GB of VRAM). It trains on a single stream at batch 1, with the attention/embedding trunk learning rate dampened to 0.1x to mitigate catastrophic forgetting.

The training is still going and went only through ~35% of the training corpus selected at the beginning of the run (about 8GB in total). Here is the current status:

1. enwik9 (First 10^9 bytes of English Wikipedia):

    Model               Params   Data Read       Cost     enwik9 (bits/byte) ↓

    Llama-3.2-1B        1.23B    36 TB (9T tok)  ~$900k   0.68
    Chinchilla 1B       1.00B    5.6 TB          ~$68k    0.90
    GPT-2 XL            1.56B    40 GB           ~$43k    1.07
    GPT-2 Large         0.77B    40 GB           ~$21k    1.12
    mini-AGI (online)   1.26B    2.8 GB          ~$56     1.14
    GPT-2 Medium        0.35B    40 GB           ~$10k    1.17
    mini-AGI (dynamic)  1.26B    2.8 GB          ~$56     1.23
    mini-AGI (frozen)   1.26B    2.8 GB          ~$56     1.29
    GPT-2 Small         0.12B    40 GB           ~$3k     1.33

2. PG19 (100 Project Gutenberg books, 41 MB test split):

    Model               Params   Trained On        Cost    PG19 (bits/byte) ↓

    MambaByte-972M       0.97B   75 GB (PG19)      ~$4k    0.883
    MegaByte-1B          1.02B   80 GB (PG19)      ~$1.6k  1.000
    Transformer-320M     0.32B   80 GB (PG19)      ~$1.6k  1.057
    mini-AGI (online)    1.26B   2.8 GB (0 PG19)   ~$56    1.447
    mini-AGI (frozen)    1.26B   2.8 GB (0 PG19)   ~$56    1.995

    The published models were trained on PG19's own training books, 75 to 80 GB of them, and are scored on more books like those. mini-AGI has never been trained on PG19 and has read no Victorian novels, so for it these books are completely out of its domain.

There are more details in the repository. The most important part is a difference between "frozen" and "online" attempts where we allow the model to specialize on the task as it goes through the evaluation dataset. Clearly showcasing the power of continual learning regime simply unavailable for any other typical LLM.

Full methodology, and replication code is in the repo: https://github.com/volotat/mini-AGI

Intermediate weights snapshot on Hugging Face: https://huggingface.co/Volotat/mini-AGI-undertrained