4 days ago
-
whisper candidate
-
whisper candidate
-
whisper candidate
-
whisper candidate
-
whisper candidate
-
whisper candidate
-
whisper probe: 6 low-voicing candidates
-
whisper probe: 6 low-voicing candidates
-
whisper probe: 6 low-voicing candidates
-
radio notebook: verify checkpoint length
-
full 4000-step student (replaces truncated upload)
-
radio demo: 4000-step student + end-to-end script
-
radio demo: 4000-step student + end-to-end script
-
radio demo: 4000-step student + end-to-end script
-
voc: stop at the ISTFT head projection, ship spec
-
fix lm cache order and additive mask; probe audio VAE
-
export: drop removed kwarg, host-side numpy
-
export: put dummy tensors on the model device
-
browser export: lm/aggregator/voc with validation + int4
-
browser export: lm/aggregator/voc with validation + int4
-
add structure probe for the browser export
-
add structure probe for the browser export
-
release candidate: 4000 steps, seeded teacher, keep audio
-
release candidate: 4000 steps, seeded teacher, keep audio
-
run4: unpack cached traj, seed teacher, 8000 steps
-
run4: unpack cached traj, seed teacher, 8000 steps
-
720 trajectory tensors, 6 voices, a07 held out (178MB)
-
run 3: 4000 steps, keep traj cache
-
holdout run: exclude a07 from training, add int4 eval stage
-
holdout run: exclude a07 from training, add int4 eval stage
5 days ago
-
ar_generate: sample under no_grad, detach latents (fixes leak)
-
eval: load student after freeing teacher, a08 only
-
eval: load student after freeing teacher, a08 only
-
surface stage failures instead of printing rc
-
eval in two passes to fit a 15GB T4
-
finish: look for student4.pt in the bundle
-
4-step student trained on 7 voices x 120 texts
-
add onnx_ir/onnxscript deps and finish notebook
-
multi-voice distillation: train_texts, train_voices.py, notebook_train.py
-
add notebook_clone.py
-
8 mined voices + clone_voices.py
-
add notebook_mining.py
-
add voice_mining.py
-
drop initial_prompt (rejected by transformers 4.46 generate kwargs)
-
fix deps: install requirement set, self-heal imports, add smoke stage
-
add notebook_eval bootstrap target
-
webfm eval set: 40 sentences, 8 voice prompts, scoring harness
-
initial commit
