Meta-Okapi/zh_bloom1b7_judgerm_muf2k_lora_decay1e-6_lr1.67e-5_500steps_seq1024

zh judge reward model (Multilingual-UltraFeedback, LoRA)

Sequence-classification reward model for zh, LoRA on bigscience/bloom-1b7. The LoRA adapter includes modules_to_save: ["classifier","score"], so the value head ships with it.

NOT comparable to the other Meta-Okapi judges

The Meta-Okapi/<lang>_bloom1b7_judgerm_decay1e-6_lr5e-5_10ksteps models are full fine-tunes on ~126,000 Okapi ranking pairs. This model is trained on a different, far smaller corpus and should not be treated as a drop-in equivalent.

Training

Multilingual-UltraFeedback (nthakur/multilingual-ultrafeedback-binarized-dpo-v0.1), from a strict id partition: the 2,200 pairs appear in no evaluation, adaptation or meta-training split in any language.

lr 1.67e-5 (constant), 500 steps, batch 16, weight_decay 1e-6, seq_length 1024, size_valid_set 0.30, eval split 300, early stopping patience 4 on eval_loss

Measured behaviour — read before using

• —Held-out pairwise accuracy: 0.611 (best checkpoint 0.623). Accuracy was still rising at step 500, so this model is underfit.
• —Length preference: r(reward, completion length) = +0.139, measured on 1,000 fixed generations. A positive value means the model scores longer completions higher independently of quality.

With 2,200 pairs seen ~5 times each, this judge has limited opportunity to learn content discrimination and measurable incentive to use surface features. Do not use it as a standalone quality metric. It was trained to make a within-experiment comparison possible, not to be a general-purpose reward model.

Scoring

Score prompt + completion with truncation_side="right" and max_length=1024, matching training. Left-truncation removes the prompt and collapses the model onto surface features.