-
docs: correct independent parameter count
-
docs: index corrected claim checkpoint timeline
Point model card to coverage-aware immutable claim revision. -
fix: report AIR coverage-aware claim metrics
Disambiguate historical valid-only AIR-Bench accuracy from coverage-aware end-to-end accuracy; checkpoint tensors are unchanged. -
docs: index paper claim checkpoint timeline
Point the model card to the immutable four-checkpoint active-interface release and explain reconstruction from the historical frozen expert export. -
fix: clarify active-interface loading and exclusions
-
release: add paper claim checkpoint timeline
-
Correct 2026-08-03 release checkpoint manifest
Updates LLM-fixed and from-phase1 checkpoint paths after verifying the Alalm-model-baseline-training worktree and uploading the checkpoint files. -
Upload AudioAlign priority no-leak from-phase1 235k checkpoint
Raw PyTorch checkpoint for sv_frozen_v3_phase2_priority_noleak_from_phase1 model-235000.pt. -
Upload Whisper priority no-leak LLM-fixed 470k checkpoint
Raw PyTorch checkpoint for whisper_largev3_fullaudio_sift_priority_noleak_llmfix model-470000.pt. -
Upload AudioAlign priority no-leak scratch 235k checkpoint
Raw PyTorch checkpoint for sv_frozen_v3_phase2_priority_noleak_scratch model-235000.pt. -
Upload Whisper priority no-leak 470k checkpoint
Raw PyTorch checkpoint for whisper_largev3_fullaudio_sift_priority_noleak model-470000.pt. -
Add 2026-08-03 priority no-leak release metadata
Adds release manifest, configs, and evaluation summaries for recent Whisper baselines and AudioAlign from_phase1/from_scratch priority no-leak experiments.
-
Upload 1.5B branch checkpoint without frozen LLM
-
initial commit
