SupraLabs' Tiny Chat Title Model Just Got a 10x Dataset Upgrade — FFT Preview Now Handles Even the Niche Topics
115K samples power SupraLabs' new fine-tuned title generator – 10x bigger than before.
SupraLabs has unveiled supra-title-FFT-preview, a specialized chat title generation model that represents a significant leap over its predecessor. The new model is trained on a carefully filtered dataset of 115K samples — roughly 10x the 12K samples used for Supra-Title-350M-exp. This expansion dramatically improves performance on uncommon conversation patterns while maintaining accuracy on common topics. The model is built on LiquidAI's LFM2.5-350M-Base (0.4B parameters) and uses full fine-tuning (FFT) rather than parameter-efficient methods like LoRA. It is a single-purpose model: given a user message, it returns a concise, relevant chat title.
Developers can integrate the model easily via Hugging Face Transformers with just a few lines of code — no system prompt required. It also supports vLLM for OpenAI-compatible serving. The model is released under Apache 2.0, making it fully open for commercial and research use. SupraLabs is actively soliciting feedback on edge cases and weird titles before locking the final, non-preview release. This checkpoint is ideal for anyone building chat-based UIs, support agents, or conversation summarization tools that need fast, lightweight title generation.
- Trained on 115K filtered samples vs previous 12K, offering 10x better coverage of niche topics.
- Uses full fine-tuning (FFT) on LiquidAI LFM2.5-350M-Base, not LoRA, for higher quality.
- No system prompt needed; generates titles directly from user messages in a few lines of code.
Why It Matters
Specialized chat title generation at 0.4B parameters with broader topic coverage and no prompt engineering needed.