v0.2.28

• Fix local_files_only flag handling by @buzzCraft in #359 (#359)
• Fix multi-task model inference by @Ingvarstep in #364 (#364)
• Add PolyLoRA support by @Ingvarstep in #365 (#365)
• Fix Docker serving-related issues by @newtondotcom in #367 (#367)
• Update serving documentation by @Ingvarstep in #368 (#368)
• Make the README more informative with a focus on the project itself by @Ingvarstep in #347 (#347)
• Update README by @urchade in #369 (#369)
• Fix token_lengths leaking into the labels encoder by @yuriihavrylko in #372 (#372)
• Preserve CLS/SEP tokens for compatible tokenizers on Transformers v5 by @yuriihavrylko in #373 (#373)
• Fix model training and saving by @Ingvarstep in #376 (#376)
• Add streaming inference support by @Ingvarstep in #377 (#377)
• Fail fast when class_token_index and vocab_size are inconsistent by @Toadoum in #371 (#371)