Teaching the model to answer.

Current phase: supervised fine-tuning

steinshark_fresh32_30b_sft has completed three SFT epochs and is now serving in the model playground.

SFT validation loss: 2.0380 · Training loss: 1.7445 · Context: 2,048 · Tokenizer: steinshark_32k.

The charts below preserve the completed 30B-token pretraining run. SFT losses use a different dataset and assistant-only targets, so they are not directly comparable to pretraining loss.