EmmaLeonhart/reservoir-agent-gpt2-batch-n20

Reservoir Agent batch — gpt2

A batch of 20 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_4.

Population

rank	seed	recall	loss_end	pr_frac	recommended
0	seed_4	1.00	0.029	0.114	yes
1	seed_9	1.00	0.283	0.113
2	seed_0	1.00	0.474	0.114
3	seed_3	1.00	0.981	0.112
4	seed_14	0.83	2.768	0.114
5	seed_19	0.67	0.037	0.113
6	seed_11	0.67	1.066	0.112
7	seed_17	0.50	0.059	0.113
8	seed_13	0.50	0.511	0.112
9	seed_2	0.50	0.783	0.113
10	seed_5	0.50	2.126	0.114
11	seed_1	0.33	0.585	0.114
12	seed_7	0.33	1.156	0.114
13	seed_18	0.33	2.479	0.112
14	seed_10	0.33	3.357	0.112
15	seed_15	0.17	1.228	0.113
16	seed_12	0.17	1.234	0.114
17	seed_16	0.17	1.614	0.114
18	seed_6	0.17	2.219	0.111
19	seed_8	0.17	2.290	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-batch-n20")
lm = load_reservoir_model(f"{path}/seed_4")

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