graf/science_gemma_1b_mix_bt_qwen4b-bt-0091fc01-1-142-on

Qwen3-4B BT — Gemma science data

Qwen3-4B sequence-scoring checkpoint trained in BonVoyage with a BT pairwise objective.

Training details

• —Base model: Qwen/Qwen3-4B
• —Training data: graf/gemma_1b_it_science_mix_train
• —Validation data: graf/gemma_1b_it_science_sciknowsci_val
• —Learning rate: 1e-5
• —Final checkpoint: epoch 141 of 142
• —Output head: one scalar (num_labels=1)
• —Weights: BF16 safetensors
• —Tokenizer: the tokenizer saved with the training run; pad_token_id=151643

Experiment: science_gemma_1b_mix_bt_qwen4b-bt-0091fc01-1-142-on.