Systematic SFT for Qwen3-4B. We explore diverse dataset compositions and training recipes to benchmark and improve performance across tasks.
AI & ML interests
Pioneering the Next Era of AI with Vector Intelligence
Recent Activity
View all activity
Making Qwen3 Think in Korean with Reinforcement Learning https://arxiv.org/abs/2508.10355
Reasoning model distilled from DeepSeek-R1, enhanced with GRPO using supplementary reasoning datasets.
For more details, please visit https://github.com/dnotitia/smoothie-qwen
-
Smoothie-Qwen: Post-Hoc Smoothing to Reduce Language Bias in Multilingual LLMs
Paper • 2507.05686 • Published • 1 -
dnotitia/Smoothie-Qwen3-0.6B
Text Generation • 0.6B • Updated • 13 • 1 -
dnotitia/Smoothie-Qwen3-1.7B
Text Generation • 2B • Updated • 637 • 2 -
dnotitia/Smoothie-Qwen3-4B
Text Generation • 4B • Updated • 16 • 4
Patched versions of Qwen3 with identical weights, featuring a separated chat_template.jinja and {% generation %} tags for seamless trl assistant_only.
High-performance LLM developed by Dnotitia Inc., incorporating cutting-edge techniques for superior reasoning tasks.
8B Korean SoTA model, which is instruction-tuned by Dnotitia Inc.
For more details, please visit https://github.com/dnotitia/smoothie-qwen
-
Smoothie-Qwen: Post-Hoc Smoothing to Reduce Language Bias in Multilingual LLMs
Paper • 2507.05686 • Published • 1 -
dnotitia/Smoothie-Qwen2.5-0.5B-Instruct
Text Generation • 0.5B • Updated • 6 -
dnotitia/Smoothie-Qwen2.5-1.5B-Instruct
Text Generation • 2B • Updated • 9 • 1 -
dnotitia/Smoothie-Qwen2.5-3B-Instruct
Text Generation • 3B • Updated • 36 • 2
Systematic SFT for Qwen3-4B. We explore diverse dataset compositions and training recipes to benchmark and improve performance across tasks.
Patched versions of Qwen3 with identical weights, featuring a separated chat_template.jinja and {% generation %} tags for seamless trl assistant_only.
Making Qwen3 Think in Korean with Reinforcement Learning https://arxiv.org/abs/2508.10355
High-performance LLM developed by Dnotitia Inc., incorporating cutting-edge techniques for superior reasoning tasks.
Reasoning model distilled from DeepSeek-R1, enhanced with GRPO using supplementary reasoning datasets.
8B Korean SoTA model, which is instruction-tuned by Dnotitia Inc.
For more details, please visit https://github.com/dnotitia/smoothie-qwen
-
Smoothie-Qwen: Post-Hoc Smoothing to Reduce Language Bias in Multilingual LLMs
Paper • 2507.05686 • Published • 1 -
dnotitia/Smoothie-Qwen3-0.6B
Text Generation • 0.6B • Updated • 13 • 1 -
dnotitia/Smoothie-Qwen3-1.7B
Text Generation • 2B • Updated • 637 • 2 -
dnotitia/Smoothie-Qwen3-4B
Text Generation • 4B • Updated • 16 • 4
For more details, please visit https://github.com/dnotitia/smoothie-qwen
-
Smoothie-Qwen: Post-Hoc Smoothing to Reduce Language Bias in Multilingual LLMs
Paper • 2507.05686 • Published • 1 -
dnotitia/Smoothie-Qwen2.5-0.5B-Instruct
Text Generation • 0.5B • Updated • 6 -
dnotitia/Smoothie-Qwen2.5-1.5B-Instruct
Text Generation • 2B • Updated • 9 • 1 -
dnotitia/Smoothie-Qwen2.5-3B-Instruct
Text Generation • 3B • Updated • 36 • 2