Experiment — Curriculum Ordering Baseline (Dataset Plan Locked)

Hypothesis (Falsifiable)

A curriculum-ordered training schedule will alter training stability and/or downstream behavior relative to a shuffled control, even when total token count and model architecture are held constant.

Variables

Independent (Changed)

• Dataset ordering: curriculum (D0 then D1 then D2) vs mixed baseline.

Controlled (Fixed)

• Source corpus: Project Gutenberg plain-text .txt.
• Language filter: English only.
• Preprocessing rules (see Dataset Specification).
• Token budget distribution: D0 25% / D1 50% / D2 25%.
• Model architecture and training loop.

Explicitly Not Tested

• Licensing constraints (not a project constraint here).
• Semantic purity of categories (segmentation is coarse and heuristic).
• RLHF or preference shaping.
• Biometric or experiential data.

Dataset Specification

• Dataset: dataset_gutenberg_exp0001_v1
• Spec: docs/projects/project_001/datasets/dataset_gutenberg_exp0001_v1.md

Runs

• Control (C): D0 + D1 + D2 fully mixed from step 0.
• Curriculum (K): D0 -> D0+D1 -> D1 -> D1+D2 -> D2.
• Run logs:
  • EXP-0001-C-runlog.md
  • EXP-0001-K-runlog.md

Results Summary (1 page max)

(Write after runs complete.)

Interpretation

(Write after results summary.)

Next Actions

• Build dataset slices for D0/D1/D2 and validate token budgets.
• Execute control vs curriculum runs with fixed seeds.