mist-tg-0.3b

mist-tg-0.3b

mist-tg-0.3b generates short chat titles from a user's first message. Fine-tuned from google/byt5-small (~300M, byte-level seq2seq) on SupraLabs/chat-titles-filtered-115K.

• Production-ready for English — trained end-to-end on real English chat messages and titles, no language tag or special formatting needed.
• Also usable on Latin-script languages as a side effect of byte-level, same-alphabet transfer — not trained on non-English targets, but Latin script's byte-level overlap with English lets same-language title extraction work reasonably well in practice for many Latin-alphabet languages (see Benchmarks and Known Limitations).
• Not reliable for non-Latin scripts (CJK, Hangul, Devanagari, Ethiopic, etc.) — see Known Limitations for why, and what would need to change to fix it.

🗒️ Model Details

🏃 Usage

from transformers import AutoTokenizer, T5ForConditionalGeneration

tok = AutoTokenizer.from_pretrained("olaverse/mist-tg-0.3b")
model = T5ForConditionalGeneration.from_pretrained("olaverse/mist-tg-0.3b")

message = "My laptop keeps freezing every time I open more than five browser tabs, any idea why?"
inputs = tok(message, return_tensors="pt", truncation=True, max_length=256)
output_ids = model.generate(**inputs, max_new_tokens=32)
print(tok.decode(output_ids[0], skip_special_tokens=True))
# Laptop Freezing Impact

No prompt template or language tag required — pass the raw message directly.

📊 Benchmarks

Cross-language qualitative check (25 languages, hand-written chat-style messages on the same topic, no ground truth — informal check, not a scored benchmark): same-language title extraction worked for roughly 17-19 of 25 languages tested, concentrated almost entirely among Latin-script languages (English, French, German, Dutch, Vietnamese, Indonesian, Yoruba, Igbo, Hausa, Swahili, Zulu, Xhosa, Shona, Somali, Afrikaans). All four non-Latin-script languages tested (Japanese, Korean, Hindi, Amharic) failed to produce same-language output, instead generating unrelated English text.

Training data & licensing

Fine-tuned from google/byt5-small (Apache-2.0) on SupraLabs/chat-titles-filtered-115K. Released under Apache-2.0.

Citation

@misc{mist-tg-0.3b,
  title  = {mist-tg-0.3b},
  author = {Olaverse},
  year   = {2026},
  url    = {https://huggingface.co/olaverse/mist-tg-0.3b}
}