Avifenesh/recipe-lab-1b

recipe-lab — 1B curriculum/architecture ablation checkpoints

Best-val checkpoints, loss curves and lm-eval readouts for every 1B-scale cell of the recipe-lab campaign: a hunt for a training recipe (data order, backbone, optimizer) that raises downstream quality for a small hybrid-MoE model.

These are research artifacts, not a product: 1.3B dense / 1.3B-total 0.41B-active MoE, ~15B tokens, GPT-2 BPE, no instruction tuning and no safety work. They exist so each claim in the campaign ledger can be re-checked against the weights that produced it.

Every cell is one variable away from its control; the reasoning, the pre-registered predictions and the retractions live in RESULTS_1B.md.

Cells (8-task lm-eval mean, 1000 samples/task)

cell	stage order	eval mean	arc_easy	arc_c	winogrande	sciq	lambada
r19_V_wl	web -> math -> phil	0.5116	0.5439	0.3278	0.525	0.633	0.362
r19_V_pm	phil -> math -> web	0.508	0.5333	0.3144	0.52	0.646	0.336
V_cur	math -> phil -> web	0.5051	0.5228	0.3344	0.53	0.607	0.35
V_nophil	math -> webx -> web (philosophy ablated)	0.503	0.5596	0.3144	0.502	0.636	0.336
V_std_ext8	proportional shuffle (no staging)	0.5013	0.5263	0.301	0.52	0.634	0.345
L_cur4	math -> phil -> web	0.4998	0.5404	0.3077	0.512	0.598	0.35
L_std	proportional shuffle (no staging)	0.4974	0.5351	0.3077	0.499	0.61	0.324
V_cur4	math -> phil -> web	0.496	0.5263	0.2876	0.531	0.607	0.319
L_std_ext8	proportional shuffle (no staging)	0.4941	0.5298	0.3144	0.499	0.608	0.315
L_cur	math -> phil -> web	0.4939	0.5316	0.2977	0.524	0.593	0.325
V_std	proportional shuffle (no staging)	0.487	0.5175	0.3043	0.541	0.62	0.334
r17_V_cur_p	math -> phil -> web	0.4821	0.5053	0.2943	0.511	0.596	0.32
r16p_HV_cur	math -> phil -> web	0.4799	0.5053	0.2977	0.503	0.564	0.32

Files per cell: <cell>_best.pt (weights), <cell>_curve.json (loss curve + full config), evals_<cell>.json (per-task accuracies).