sunbi10/gemma-4-31B-it-itb-v27
The sunbi10/gemma-4-31B-it-itb-v27 is a 31 billion parameter instruction-tuned Gemma-4 model, post-trained by sunbi10 specifically for the KEPCO Korean power/plant engineering domain. It features a 32768 token context length and incorporates advanced training techniques like GRPO Math and On-Policy Distillation. This model demonstrates significant performance improvements in domain-specific benchmarks, making it highly specialized for technical applications within the power and plant engineering sector.
Loading preview...
Model Overview
The sunbi10/gemma-4-31B-it-itb-v27 is a 31 billion parameter instruction-tuned Gemma-4 model, specifically post-trained for the KEPCO Korean power/plant engineering domain. This model leverages a 32768 token context length and integrates a sophisticated training pipeline including Supervised Fine-Tuning (SFT), GRPO Math, On-Policy Distillation (OPD), and GRPO IFEval stages.
Key Capabilities & Training
- Domain Specialization: The model is explicitly fine-tuned on KEPCO Korean power/plant engineering data, enhancing its relevance and performance in this technical field.
- Advanced Training Pipeline: It utilizes a multi-stage training approach:
- SFT (v21): Incorporates kNN retrieval and synthetic data for initial fine-tuning.
- GRPO Math (v25): Further trained with Nemotron AOPS+SE and a math verification reward.
- OPD (base ← v25): Employs on-policy distillation from the v25 version.
- GRPO IFEval (v27): Refined with
allenai/RLVR-IFevalsamples and anIF_FUNCTIONS_MAPreward.
Performance Highlights
- ITB MCQ: Achieves 90.36%, a notable improvement of +5.52 over the base model in domain-specific multiple-choice questions.
- KMMLU: Scores 81.59%.
- GPQA Diamond: Reaches 81.82%.
- IFEval (strict prompt): Demonstrates 95.19% accuracy.
- Qualitative Assessment: Maintains an overall qualitative score of 4.92/5.00 with no degradation compared to the base model across various categories including ITB free-form, Korean general, and writing tasks.
- Knowledge Check: Shows improved performance in domain-specific knowledge checks, with gemma judge scoring 3.90 (vs. 2.80 base) and Claude judge scoring 3.60 (vs. 2.20 base).
Recommended Use Cases
This model is particularly well-suited for applications requiring deep understanding and generation within the Korean power and plant engineering domain. Its specialized training makes it ideal for tasks such as technical documentation, question answering, and knowledge retrieval in this specific industry.