ZaandaTeika/Qwen2.5-1.5B-SHARP-Span

Qwen2.5-1.5B-SHARP-GSM8K-proc_30-70-Math-proc_10-90

Process Reward Model (PRM) based on Qwen/Qwen2.5-Math-1.5B-Instruct.

Training

Two-stage fine-tuning:

1. proc_30-70 — GSM8K-style process supervision (SHARP)
2. proc_10-90 — Math dataset (SHARP)

Weights are stored in bfloat16.

Load

from transformers import AutoModel, AutoTokenizer

model_id = "ZaandaTeika/Qwen2.5-1.5B-SHARP-GSM8K-proc_30-70-Math-proc_10-90"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, torch_dtype="bfloat16", trust_remote_code=True)

Custom architecture: Qwen2ForProcessRewardModel (see modeling_qwen2_rm.py in the repo).

Format

Step separator: \n\n. Each completion step is labeled at its last token (0 = error, 1 = correct).