Release 0.3.0

What's new:

• Removed unused entries in C2K (upper case ASCII, they were there for a mistake).
  • It changes the embedding dim, I treat this as an API change and bump the minor version.
• Use a slightly larger training split.
• Add ' into input symbols for C2K and P2K, this boosted the bleu score a little bit.

This release contains following resources:

• model-{c2k/p2k}-e10.pth: raw state_dict files from pretrained models in PyTorch, fp32, 10 epochs
• model-{c2k/p2k}.npz: npz files directly used in production library, fp16
• model-{c2k/p2k}.safetensors: safetensors files for integration in other languages, fp16

Resources didn't change from 0.2.0 is not included:

• katakana_dict.jsonl: a JSON line file contains the extracted English to Katakana dictionary

Check 0.2.0 for these resources.

Please acknowledge that the dictionary should follow the Wikimedia's license and the JMdict / EDICT's Copyright license.

I'm not a lawyer, how to treat the model weights derived from the dictionary is up to you.