Modelpublic

abugoot-primeintellect/bioreasoning-qwen3-30ba3b-sft-20260218-ckpt-final

sourceHugging Faceupdated 7mo agoView on Hugging Face

0likes8downloads

Model Card

bioreasoning-qwen3-30ba3b-sft-20260218-ckpt-final

LoRA adapter checkpoint from supervised fine-tuning on the bioreasoning dataset.

—Base model: Qwen/Qwen3-30B-A3B
—Method: LoRA (rank 32)
—Training platform: Tinker (ThinkingMachines)
—Dataset: abugoot-primeintellect/bioreasoningv0211prime_sft (~193k train examples)
—Hyperparameters: batchsize=128, lr=2e-5 (linear decay), weightdecay=0.01, seq_len=8192
—Epochs: 3
—Checkpoint: step final (epoch 3, batch 1511)
—Loss masking: Last assistant message only

Step	Epoch	HF Repo
1500	~1	bioreasoning-qwen3-30ba3b-sft-20260218-ckpt-1500
3000	~2	bioreasoning-qwen3-30ba3b-sft-20260218-ckpt-3000
final	3	bioreasoning-qwen3-30ba3b-sft-20260218-ckpt-final