diff options
| author | YurenHao0426 <blackhao0426@gmail.com> | 2026-06-13 12:35:36 -0500 |
|---|---|---|
| committer | YurenHao0426 <blackhao0426@gmail.com> | 2026-06-13 12:35:36 -0500 |
| commit | 66e0d8b9fd4d0f7a2231d689c055e26fdf1cf04a (patch) | |
| tree | c29cba61124018755a19b02c9d33e3ad5f2e05cc /research/flossing/report_bundle_20260603/README.md | |
Curated export for clone-and-run Maze training (2x A6000) + diagnostics.
trm/hrm pretrain.py carry trajectory-augmentation code (backward-compatible).
Heavy artifacts (checkpoints/wandb/npz) gitignored; see PROVENANCE.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Diffstat (limited to 'research/flossing/report_bundle_20260603/README.md')
| -rw-r--r-- | research/flossing/report_bundle_20260603/README.md | 34 |
1 files changed, 34 insertions, 0 deletions
diff --git a/research/flossing/report_bundle_20260603/README.md b/research/flossing/report_bundle_20260603/README.md new file mode 100644 index 0000000..ea92ac5 --- /dev/null +++ b/research/flossing/report_bundle_20260603/README.md @@ -0,0 +1,34 @@ +# RRM Report Bundle 2026-06-03 + +This bundle contains report/PPT-ready artifacts for the recursive reasoning dynamics work. + +## Recommended Figures + +- `figures/fig0_motivation_lambda1_success_failure_hrm_trm.png`: motivation, first Lyapunov exponent separates success and failure for HRM/TRM. +- `figures/fig1_hrm_trm_training_curves.png`: HRM/TRM baseline versus multi4 training curves. +- `figures/fig2_accuracy_vs_chaotic_volume_phase.png`: accuracy versus chaotic volume view. +- `figures/fig3_hrm_trm_success_failure_spectra.png`: success/failure full-spectrum separation. +- `figures/fig4_ptrm_same_subset_comparison.png`: PTRM same-subset comparison. +- `figures/fig5_qhead_vs_lambda1_ptrm.png`: PTRM Q-head score versus finite-difference stability proxy. + +## Tables + +- `tables/meeting_figures_v2_report.md`: concise figure strategy and caveats. +- `tables/hrm_trm_redesigned_summary.csv`: HRM/TRM headline accuracy and spectrum metrics. +- `tables/fig5_qhead_vs_lambda1_ptrm_summary.csv`: Q-head versus stability statistics. +- `tables/*eval*.csv`: eval curves and checkpoint comparisons used to generate the figures. +- `tables/*summary*.csv`: PTRM selection and Lyapunov diagnostic summaries. + +## Raw Data + +- `raw_npz/`: compact diagnostic arrays backing the main figures. +- Excludes model checkpoints, W&B full history dumps, and very large tangent-mode dumps. + +## Scripts + +- `scripts/make_meeting_artifacts_v2.py`: regenerates the main HRM/TRM/PTRM figure set. +- `scripts/make_q_lambda_scatter.py`: regenerates the Q-head versus stability figure. + +## Supplemental + +- `supplemental_figures/`: earlier exploratory PNGs that may be useful as backup slides. |