Mun2/qwen3.5-9b-korean-essay-scorer-vllm
The Mun2/qwen3.5-9b-korean-essay-scorer-vllm is a 9 billion parameter model based on Qwen3.5-9B, specifically fine-tuned for scoring Korean argumentative essays. It generates JSON output for 'content', 'organization', and 'expression' scores (1-5) along with Korean rationales. This model is optimized for automated essay evaluation, providing structured feedback for Korean text analysis.
Loading preview...
Overview
This model, developed by Mun2, is a specialized Qwen3.5-9B variant (9 billion parameters, 4096 token context length) designed for automated scoring of Korean argumentative essays. It integrates a GRPO LoRA into the base Qwen3.5-9B weights, allowing it to be loaded directly with standard Transformers and vLLM without custom code or adapters.
Key Capabilities
- Korean Essay Scoring: Provides scores (1-5) for
content,organization, andexpressionof Korean argumentative essays. - Rationale Generation: Accompanies each score with a Korean rationale explaining the judgment.
- JSON Output: Generates structured JSON output, making it easy to parse and integrate into applications.
- Optimized for Evaluation: Trained with a fixed chat template and specific prompt/essay input format for essay scoring tasks.
Performance & Limitations
Validation results show 97.5% valid JSON output, with a Macro integer RMSE of 0.95881 and Macro integer Spearman of 0.44608 on 390 valid samples. While effective for automated assessment, the model's output should not be used as the sole basis for high-stakes decisions due to a small percentage of format failures and the inherent nature of automated scoring.