SPB Git

spb/wp31_uqo

Public
971 B · 32 lines yaml
1# Locked evaluation of the 103M pilot: held-out token cross-entropy at forced budgets and under the controller.2# Splits: val (development) now; test (LOCKED) only after protocol/preregistration.md is frozen and the manifest unsealed.3out_dir: results/raw4data: {kind: tokens, dir: data/processed/fineweb_edu_300m}5splits: [val]6seq_len: 5127batch_size: 168n_batches: 40          # 40 x 16 x 512 = 327,680 held-out tokens9budgets: [0, 1, 2, 4, 8]10systems:11  backbone_103m:12    mode: backbone_only13    model: from_checkpoint14    checkpoint: PRETRAIN103M_CKPT15    controller: false16  enlarged_102m:17    mode: backbone_only18    model: from_checkpoint19    checkpoint: ENLARGED_CKPT20    controller: false21  prisme_103m_editor:22    mode: prisme23    model: from_checkpoint24    checkpoint: EDITOR103M_CKPT25    controller: false26  gr_103m:27    mode: global_recurrence28    baseline_kwargs: {ffn_w: 2560}29    model: from_checkpoint30    checkpoint: GR103M_CKPT31    controller: false32
Files · 20e88d79efe28fb13610284e7be1ef0ef40c265d
Menu

Theme