Bc-AI/T1-Mini-Preview
SmilyAI Labs T1-Mini-Preview
T1-Mini-Preview is an early preview of the upcoming T1-Mini model from SmilyAI Labs.
T1-Mini-Preview is based on Qwen/Qwen3.5-4B and was fully fine-tuned on Bc-AI/SFT-Ultra, a curated instruction-tuning dataset prepared for SmilyAI's small-model research.
Training
The model was trained in two main stages:
- Supervised Fine-Tuning (SFT) — trained the base model on our curated instruction and reasoning data.
- Direct Preference Optimization (DPO) — further refined the model's responses using preference-based training.
This two-stage pipeline was designed to improve instruction following, response quality, and overall conversational behavior while keeping the model relatively small and efficient.
Model Status
T1-Mini-Preview is a preview release, not the final T1-Mini model. Training, evaluation, and further refinement are still ongoing.
We are releasing this version so the community can experiment with it and provide feedback while development continues.
Base Model
- Base: Qwen/Qwen3.5-4B
- Training: Full fine-tuning
- Primary language: English
- Fine-tuning dataset: Bc-AI/SFT-Ultra
- Training stages: SFT → DPO
- License: Apache 2.0
About SmilyAI Labs
SmilyAI Labs is a small open-source AI project focused on building capable, efficient, and accessible AI models.
We're experimenting with smaller models that can deliver strong performance without requiring enormous amounts of compute.
🚀 T1-Mini-Preview is one step toward that goal.
Disclaimer
This is an experimental preview model. Its behavior and capabilities may differ from the final T1-Mini release, and it may occasionally produce incorrect, inconsistent, or undesirable outputs.