Dataset Viewer
Auto-converted to Parquet Duplicate
status
string
commit
string
runner_sha
string
plan_sha
string
total_runtime_s
float64
mimi_device
string
four_cell_cb0
dict
g2_diff_in_diffs_D
float64
g2_survival_fraction
float64
interpretation
string
C7_RETRAIN_VCTK_LORA_ADAPTATION_COMPLETE
66066ca
d8013be6e9d03a26194a72d8f3b098e8d826bdb7a4f1f915e578a934eba8956e
1fb2930c4e7b061b4718794f715627dfc0db851fa5e40f8ca80da2248ea5f0c1
5,971.323
cpu
{ "control_seen": 0.649775, "control_novel": 0.663079, "lora_seen": 0.202356, "lora_novel": 0.202621 }
0.013039
1.029143
D > 0 meaningfully + survival ≈ 1: LoRA gain is genuine content-disjoint generalization, not just content-recognition

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

C7 VCTK Retrain

Seam-3 C7-retrain VCTK content-disjoint evaluation.

Design

  • Retrain from scratch on content-reserved train_4000 (40 text_ids R reserved for novel eval)
  • Two arms within run:
    1. control_no_lora: LoRA attached but frozen, AF+DT trainable
    2. lora_trainable: q/v r16 LoRA trainable, AF+DT trainable
  • Eval on same 25 heldout speakers with:
    • content_seen: 512 windows, text_ids NOT in R (seen during training)
    • content_novel: 512 windows, text_ids IN R (unseen during training)
  • Primary metric: diff-in-diffs D = (control_novel − lora_novel) − (control_seen − lora_seen)

Results

Cell cb0_real_over_baseline
control_seen 0.650
control_novel 0.663
lora_seen 0.202
lora_novel 0.203
D 0.013 (positive)
Survival 1.029

Interpretation

D > 0 and survival ≈ 1: LoRA gain genuinely generalizes to novel (never-trained) text_ids. Not just content-recognition. The C6 LoRA effect survives within-VCTK content-disjoint eval.

Artifacts

  • c7_retrain_essential.tar.gz: report.json, report.md, checkpoints/
Downloads last month
19