safetensors + config for the v2 checkpoints
#1
by phas02 - opened
Adds safetensors conversions of the two v2 checkpoints, plus a small
config.json per checkpoint, so consumers outside torch (Rust/candle,
safetensors readers generally) can fetch the weights without loading a
pickle:
- tabicl-classifier-v2-20260212.safetensors / .config.json
- tabicl-regressor-v2-20260212.safetensors / .config.json
The conversion is mechanical: every tensor in the checkpoint's
state_dict saved as-is (all plain float32, no shared storage), and the
checkpoint's own config dict written as JSON alongside. Nothing is
renamed or transformed, so state_dict consumers and safetensors
consumers see identical tensors.
Verified against the torch forward: a candle port of the v2
architecture loads exactly these files and matches the torch
train-mode forward at ~2e-5 max abs diff (fp32) on both checkpoints,
CPU and Metal.