v5 codec: v5-onnx-manifest.json

v5 codec: v5-encoder-int8.onnx

v5 codec: v5-encoder-fp32.onnx

v5 codec: v5-encoder-fp16.onnx

v5 codec: v5-decoder-int8.onnx

v5 codec: v5-decoder-grad-fp32.onnx

v5 codec: v5-decoder-fp32.onnx

v5 codec: v5-decoder-fp16.onnx

v5 codec: v5.pt

v4 checkpoint: cc12 lineage, stage-2 fine-tune through the PROTOCOL_VERSION 3 modem, epoch 536

v4 codec artifacts (from v4.pt, PROTOCOL_VERSION 3 modem fine-tune, epoch 536)

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v3 (cc12) ONNX artifacts + decoder gradient graph

v2 checkpoint: s2 + nonphoto-frac 0.1 fine-tune (epoch 317)

ONNX artifacts for v2.pt

ONNX artifacts for v2.pt

ONNX artifacts for v2.pt

ONNX artifacts for v2.pt

ONNX artifacts for v2.pt

ONNX artifacts for v2.pt

ONNX artifacts for v2.pt

export: measured per-layer int8 tuning against off-distribution probes

model card: tuned int8 figures

int8: keep each part's most quantisation-sensitive layer at fp32, tuned against off-distribution probes

re-export: int8 encoder keeps its most quantisation-sensitive layer at fp32 (5.8 -> 14.0 dB under channel noise)

model card: document the ONNX artifacts

publish-time ONNX export script

ONNX artifacts for v1.pt

ONNX artifacts for v1.pt

ONNX artifacts for v1.pt

ONNX artifacts for v1.pt

ONNX artifacts for v1.pt

ONNX artifacts for v1.pt

ONNX artifacts for v1.pt

link the project repository

add model card

default to the published checkpoint when --model is omitted

keep only model/training code (match sstvae-code minus train_job.py)

add project source, CLI tools and tests (no train_job.py)

Upload v1 checkpoint

initial commit