The encoder/classifier are rebuilt in bf16/fp32 from the code-free
model.safetensors via src/models/v2_inference.py, bit-identical to the
removed graphs. TorchScript is a code container that model scanners flag
regardless of content (Protect AI PAIT-TCHST-301).
* remove classifier.ts
* remove classifier_gpu.ts
* remove encoder.ts
* remove encoder_gpu.ts
* remove torchscript_parity.json
* .gitattributes
* README.md
* centroid_report.json
* config.json
* extension/moe/CMakeLists.txt
* plant_v2_cli.py
* requirements.txt
* src/models/v2_inference.py
* SHA256SUMS.json
20 days ago
Update: single-source model card (consistent V2 title, custom-kernel section, PTE run instructions)
* reported scores divided by a fitted temperature of 1.90 (held-out ECE 0.101 -> 0.062, grouped 5-fold CV); top-1/top-5 and all argmax-based checks are invariant
* --tta {5,10,12} fuses crops of a single photo in probability space (top-1 86.33% -> 87.05%, top-5 93.47% -> 94.62%)
* README: Harness measurements section with the benchmark tables and CPU cost
* sync the already-committed expert widen fix in src/ (no-op at widen=1.0)
22 days ago
Publish the best V2 mean-pooling model with the W8A8 ExecuTorch PTE and AdaRound-FP8 weights
