Developer Shares Post-Training Challenges Distilling Chain-of-Thought into AliceAI-80B

An open-source developer shared progress on locally fine-tuning Yandex's AliceAI-80B-A3B model using three 32GB Nvidia V100 GPUs. After an initial round of supervised fine-tuning (SFT) on 5 million tokens resulted in severe underfitting for general conversation, the developer expanded synthetic data generation using their open-source tool, sftmill. This practical experiment demonstrates both the feasibility and token-breadth requirements of post-training modern Mixture-of-Experts (MoE) models on older enterprise hardware. It highlights how accessible off-policy distillation engines enable individual developers to teach open models specialized reasoning capabilities. The first checkpoint successfully learned basic chain-of-thought reasoning but produced garbled outputs on ambiguous queries due to a narrow dataset focused primarily on coding. To fix this, synthetic data generation throughput was tripled to 240 tokens per second across multiple parallel instances, preparing 5 million broader general-instruction tokens for a second training pass.