hemisphere-llm/hemispheres-step1

Hemispheres: step-1 checkpoints

Weights for the step-1 synthetic-world runs of Hemispheres, a language model that keeps its knowledge in a separate, editable store and reads it inside the forward pass. Each run directory holds config.json (every training flag, seed and model shape) and MLX safetensors checkpoints.

The run records (metrics, per-question evaluation outputs, data fingerprints), the comparison report and the code to load and evaluate these weights are in the project repository, under results/step1/.

run	arm	params	checkpoints	sha256
context-a	context	25.5M	final, latest	ae48aaa6bc50
dense-a	dense	25.5M	final, latest	81e8eb4ac912
dense-a-k100	dense	25.5M	final, latest	c2535e281e08
dense-a-to-b	dense	25.5M	final, latest	17c630e69033
latent-a	latent	29.2M	final, latest	f4f3b1fbe96a
latent-a-all	latent	29.2M	final, latest	57dc7d12714c
latent-multi	latent	29.2M	final, latest	1b2490e51d93
latent-multi-nohop	latent	29.2M	latest	dc2494c38336
lookup-a	lookup	25.5M	final, latest	e06385a28712

pip install -e '.[hub]'
python -m hemispheres.records fetch latent-multi          # into runs/latent-multi/, sha256-checked
python -m hemispheres.synth.build --name world-b --index 1 --edits 1,100,1000
python -m hemispheres.evaluate --run runs/latent-multi --data data/world-b --sets all --n 300

Models are 25.5M (dense, lookup, context) or 29.2M (latent) parameters, fp32, trained from scratch on generated worlds. They are research artifacts for the synthetic task, not general-purpose language models.

Released under the MIT license (see LICENSE).