EmmaLeonhart/reservoir-agent-gpt2-medium-batch

Reservoir Agent batch — gpt2-medium

A batch of 10 reservoir agents (different fixed-random reservoir seeds) trained on the cross-pass recall task. A reservoir agent is a new model type: a pretrained transformer with a fixed reservoir brain-surgeried in (attended, cross-pass-stateful, RNN-like) — see the project and RESERVOIR_AGENTS.md.

The whole population is published, not just the winner. Reservoir performance is stochastic in the seed; the suboptimal models are kept as signal for learning which reservoir properties survive selection. The recommended model is seed_9.

Population

rank	seed	recall	loss_end	pr_frac	recommended
0	seed_9	0.17	1.269	0.113	yes
1	seed_1	0.17	1.362	0.114
2	seed_3	0.17	1.625	0.112
3	seed_8	0.17	1.683	0.114
4	seed_5	0.17	1.738	0.114
5	seed_2	0.17	1.902	0.113
6	seed_6	0.17	2.283	0.111
7	seed_4	0.17	2.345	0.114
8	seed_7	0.17	2.442	0.114
9	seed_0	0.17	2.524	0.114

Use

Each seed_<n>/ is a complete loadable reservoir agent. Load the recommended one:

from huggingface_hub import snapshot_download
from reservoir.persist import load_reservoir_model
path = snapshot_download("EmmaLeonhart/reservoir-agent-gpt2-medium-batch")
lm = load_reservoir_model(f"{path}/seed_9")

batch_manifest.json records the ranking + each seed's score and reservoir-dynamics signal.