Persisted experiment outputs. Small derived files (CSV/JSON summaries, bootstrap
results, training reports) are tracked in git so the paper and figures are reproducible
offline. Large raw files (N-best and PLL JSONLs, 10–140 MB each) are not tracked —
they live on Google Drive (see the README Data availability section). All *.jsonl
are git-ignored by default.
| File | Backs | Contents |
|---|---|---|
gap_covering/stage2/gamma_summary.json |
Fig. 2 | CTC alignment-posterior γ_t three-way split (dead/active/ambiguous) |
exp_A_mwer/smoke_report.json, exp_B_mwer_clipped/smoke_report.json |
Fig. 3a | CR-CTC MWER training trajectories (unclipped / clipped) |
standard_ctc/mwer_smoke_report.json |
Fig. 3b | Standard-CTC MWER mild-drift trajectory |
g_scaling/scaling_curve.csv |
Fig. 4 | WER vs beam size G (greedy/oracle/interp/MBR) |
tau_sweep/tau_sweep.csv |
Fig. 5 | WER vs MBR temperature τ at G=128, with bootstrap CIs |
g_scaling/scaling_spearman.csv |
Fig. 6 + teaser | Per-scorer Spearman ρ vs G (CTC / PLL / GPT-2 / interp) |
e25_bootstrap/gap_e25_bootstrap.json |
Fig. 7 | Cross-condition WER-change + bootstrap significance |
Regenerate all figures with python scripts/generate_figures.py (or bash scripts/reproduce_figures.sh).
dev_clean/,test_other/,test_other_g128/,zipformer_m*/,voxpopuli/,tl3_*,musan_*— per-condition decoding results and bootstrap CIs behind the §6 master table.significance/,e25_bootstrap/— paired-bootstrap significance (B=10,000, seed 42).g_scaling/scaling_bootstrap.csv,optimal_alpha_per_G.csv— beam-scaling diagnostics.diagnostics/,regression_analysis/,prop_verification/,S3_prop1_g128/,R_*— proposition checks, seed-variance, and intervention runs.
**/nbest_*.jsonl,**/neural_lm_scores_*.jsonl— raw N-best candidates and PLL/LM scores.- Model checkpoints (
*.pt), audio (*.wav/*.flac), and the CER-matrix cache (results/.cache/).