dongboklee/gORM-14B-merged
TEXT GENERATIONConcurrency Cost:1Model Size:14.8BQuant:FP8Ctx Length:32kPublished:Sep 29, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold
The dongboklee/gORM-14B-merged model is a 14.8 billion parameter LoRA-merged version of gORM-14B, developed by Dong Bok Lee and collaborators. This model is specifically designed for vLLM inference and is based on research into rethinking reward models for multi-domain test-time scaling. It is optimized for tasks requiring detailed solution verification and grading, particularly in educational or technical contexts, leveraging its 131072 token context length.
Loading preview...
gORM-14B-merged Overview
The dongboklee/gORM-14B-merged is a 14.8 billion parameter language model, representing a LoRA-merged version of the original gORM-14B. Developed by Dong Bok Lee and his team, this model is primarily intended for efficient inference using vLLM.
Key Capabilities
- Reward Model Specialization: This model is rooted in research focused on "Rethinking Reward Models for Multi-Domain Test-Time Scaling," suggesting its strength in evaluating and grading responses.
- Solution Verification: The provided example demonstrates its use as a "teacher" to grade solutions, verifying correctness step-by-step and providing a final 'Yes/No' assessment.
- Extended Context Length: With a context length of 131072 tokens, it can process and analyze extensive problem descriptions and solutions.
Good For
- Automated Grading: Ideal for applications requiring automated assessment and verification of answers, particularly in technical or educational domains like computer science.
- Detailed Feedback Generation: Capable of generating step-by-step verification processes for solutions.
- Research in Reward Models: Useful for researchers exploring multi-domain reward model scaling and inference efficiency.