friendshipkim/Qwen2.5-Math-1.5B-Scoring-Mean

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kPublished:Nov 16, 2025License:apache-2.0Architecture:Transformer Open Weights Warm

The friendshipkim/Qwen2.5-Math-1.5B-Scoring-Mean is a 1.5 billion parameter Qwen2-based model with a 131072 token context length, featuring a unique dual-head architecture. In addition to standard next-token prediction, it includes a dedicated 'Success Rate Head' designed to predict a probability score for the generated sequence. This model is specifically engineered for tasks requiring both text generation and an assessment of the output's success or correctness, particularly in mathematical or reasoning contexts.

Loading preview...

Overview

This model, friendshipkim/Qwen2.5-Math-1.5B-Scoring-Mean, is a specialized 1.5 billion parameter Qwen2-based language model with a 131072 token context window. Its core innovation lies in its dual-head architecture:

  • A standard Language Model Head for next-token prediction and text generation.
  • A unique Success Rate Head that outputs a probability score (between 0 and 1) indicating the likelihood of the generated sequence being correct or successful.

This design allows the model to not only produce answers but also to provide a confidence score for those answers, making it particularly useful for tasks where evaluating the correctness of an output is crucial.

Key Capabilities

  • Dual Output: Simultaneously generates text and provides a success probability score for the generated sequence.
  • Qwen2 Backbone: Leverages the robust Qwen2 transformer architecture for its language modeling capabilities.
  • Custom Modeling: Utilizes a custom modeling_custom.py file to integrate the scoring head, accessible via the return_score parameter during inference.

Training and Usage

The success rate head is randomly initialized and requires fine-tuning on specific tasks to learn meaningful success probabilities. The base model and LM head are initialized from friendshipkim/Qwen2.5-Math-1.5B. Developers can easily integrate this model using the transformers library, setting return_score=True to obtain both the generated text and its associated success probability.

Finetunes
1 models