You can deploy this model and start earning money today!

aviralchawla/
metaothello

Name: metaothello
Availability: InStock

pytorch

othello

mechanistic-interpretability

Views

No views yet

MetaOthello: Pretrained Models & Board Probes

Pretrained GPT checkpoints and linear board probes for MetaOthello, a controlled suite of Othello game variants used to study how transformers organize multiple world models internally.

Paper: MetaOthello: A Controlled Study of Multiple World Models in Transformers

Code: github.com/aviralchawla/metaothello

Training Data: datasets/aviralchawla/metaothello

Repository Contents

GPT Model Checkpoints

Seven trained GPT models — four single-game and three mixed-game — each with checkpoints at epochs 1, 5, 50, 150, and 250:

Directory	Training Data	Description
`classic/`	Classic (20M)	Standard Othello
`nomidflip/`	NoMidFlip (20M)	Flip endpoints only
`delflank/`	DelFlank (20M)	Delete flanked pieces
`iago/`	Iago (20M)	Classic with scrambled token vocabulary
`classic_nomidflip/`	Classic + NoMidFlip (20M each)	Mixed: high rule overlap
`classic_delflank/`	Classic + DelFlank (20M each)	Mixed: low rule overlap
`classic_iago/`	Classic + Iago (20M each)	Mixed: isomorphic control

Architecture (all models): 8 layers, d_model=512, 8 attention heads, vocabulary size 66, context window 59 tokens.

Board Probe Checkpoints

Linear probes trained to predict tile state (Mine / Opponent / Empty) from residual-stream activations at each layer. Located under board_probes/{run_name}/:

Probe directory	Probes	Description
`board_probes/classic/`	8	One per layer, trained on Classic data
`board_probes/nomidflip/`	8	One per layer, trained on NoMidFlip data
`board_probes/delflank/`	8	One per layer, trained on DelFlank data
`board_probes/iago/`	8	One per layer, trained on Iago data
`board_probes/classic_nomidflip/`	16	8 Classic + 8 NoMidFlip probes
`board_probes/classic_delflank/`	16	8 Classic + 8 DelFlank probes
`board_probes/classic_iago/`	16	8 Classic + 8 Iago probes

Naming convention: {game}_board_L{layer}.ckpt (layers 1–8).

Coming soon: Game ID probes (linear classifiers that predict which game variant is being played), along with training and plotting scripts, are currently in development and will be uploaded in a future update.

File Layout

aviralchawla/metaothello/
├── classic/
│   ├── epoch_1.ckpt          # ~101 MB each
│   ├── epoch_5.ckpt
│   ├── epoch_50.ckpt
│   ├── epoch_150.ckpt
│   └── epoch_250.ckpt
├── nomidflip/                # same structure
├── delflank/                 # same structure
├── iago/                     # same structure
├── classic_nomidflip/        # same structure
├── classic_delflank/         # same structure
├── classic_iago/             # same structure
└── board_probes/
    ├── classic/
    │   ├── classic_board_L1.ckpt    # ~396 KB each
    │   └── ...through L8
    ├── classic_nomidflip/
    │   ├── classic_board_L1.ckpt    # ...through L8
    │   └── nomidflip_board_L1.ckpt  # ...through L8
    └── ...                          # other runs

Usage

Download via the MetaOthello CLI (recommended)

bash

# Clone the repository
git clone https://github.com/aviralchawla/metaothello.git
cd metaothello && pip install -e .

# Download all pretrained assets (models + data + board probes)
make download-all

# Download all GPT models
make download-models

# Download all board probes
make download-probes

# Download selectively
make download-model RUN_NAME=classic               # Single model
make download-board-probe RUN_NAME=classic_iago     # Single run's probes

Models are placed into data/{run_name}/ckpts/ and probes into data/{run_name}/board_probes/.

Download with huggingface_hub

python

from huggingface_hub import snapshot_download

# Download a single model's final checkpoint
snapshot_download(
    repo_id="aviralchawla/metaothello",
    repo_type="model",
    allow_patterns=["classic/epoch_250.ckpt"],
    local_dir="./data",
)

# Download all board probes for a run
snapshot_download(
    repo_id="aviralchawla/metaothello",
    repo_type="model",
    allow_patterns=["board_probes/classic_nomidflip/*.ckpt"],
    local_dir="./data",
)

Load a model

python

from metaothello.mingpt.utils import load_model_from_ckpt

# Load as a minGPT model
model = load_model_from_ckpt("data/classic/ckpts/epoch_250.ckpt", vocab_size=66, block_size=59)

# Load as a TransformerLens HookedTransformer (for mechanistic interpretability)
model = load_model_from_ckpt(
    "data/classic/ckpts/epoch_250.ckpt", vocab_size=66, block_size=59, as_tlens=True
)

Training From Scratch

To retrain models or probes, see the full instructions in the GitHub README.

Note: Board probe training requires caching residual-stream activations locally (>1 TB total across all models), which is why only the trained probe checkpoints (~396 KB each) are hosted here rather than the cached activations. Pre-trained probes are sufficient for all analysis scripts in the repository.

Citation

bibtex

@article{metaothello2025,
  title   = {MetaOthello: A Controlled Study of Multiple World Models in Transformers},
  author  = {Aviral Chawla, Galen Hall, Juniper Lovato},
  journal = {arXiv preprint},
  year    = {2025}
}

License

MIT