Search papers, labs, and topics across Lattice.
Alibaba Group
1
0
0
3
Failure-driven post-training, combined with a meticulously curated 10M token STEM dataset, unlocks a 4.68% performance boost in LLM reasoning, proving that strategic data synthesis around model weaknesses is a powerful path to improvement.
