EmmaLeonhart/reservoir-agent-gpt2-batch-n12

Reservoir Agent batch — gpt2

A batch of 12 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_1.

Population

rank	seed	recall	loss_end	pr_frac	recommended
0	seed_1	1.00	0.003	0.114	yes
1	seed_7	1.00	0.005	0.114
2	seed_10	1.00	0.271	0.112
3	seed_6	0.83	0.881	0.111
4	seed_5	0.67	0.709	0.114
5	seed_4	0.67	1.280	0.114
6	seed_2	0.67	3.503	0.113
7	seed_3	0.50	0.868	0.112
8	seed_9	0.33	1.510	0.113
9	seed_0	0.33	1.975	0.114
10	seed_8	0.17	1.862	0.114
11	seed_11	0.17	2.082	0.112

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-n12")
lm = load_reservoir_model(f"{path}/seed_1")

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