Report
xLLM training library launches with an adjustable tokenizer and chat template
A developer behind xLLM says quick adjustments helped the team fix small chat-template bugs during training.
TLDR
A developer behind xLLM announced the training library’s release and highlighted its flexibility. They say the team kept finding small bugs in its chat template and could fix them during training by quickly adjusting the tokenizer and template. The final Jinja template was 1,000 lines long.
Combined views
3
1 Source, first seen 1h ago
reposts