davemaxuellkr/260814-writer-model

Overview

This model, davemaxuellkr/260814-writer-model, is a specialized Korean essay-evaluation rationale writer, fine-tuned from Qwen/Qwen3.5-9B. It is designed to generate detailed Korean rationales for argumentative essays based on pre-determined scores for content, organization, and expression. Crucially, it is not a score prediction model but rather a rationale generator, intended to be part of a larger pipeline.

Key Capabilities

• Score-Conditioned Rationale Generation: Takes integer scores (1-5) for content, organization, and expression, along with the original essay text, to produce specific Korean evaluation rationales.
• Pipeline Integration: Functions as a 'writer' in a multi-step system, where an upstream scorer first predicts essay scores, which are then fed into this model.
• Structured Output: Generates a nested JSON object containing the echoed score and the generated rationale for each category (content, organization, expression).

Performance

The model achieved a 4.3926 rationale-quality proxy score on a 400-essay validation protocol, outperforming previous SFT and DPO-B writers under the same conditions. This score reflects the quality of the generated rationales in terms of domain relevance, score-rationale consistency, specificity, and fidelity to the original text. It's important to note that this is a local judge-proxy result and not an official leaderboard score or a measure of score prediction accuracy.

Usage Considerations

• Requires an external scorer to provide initial content, organization, and expression scores.
• Utilizes specific system prompts and user templates for native inference, ensuring consistency with its training contract.
• The generated score in the output JSON is a copy of the input score; the model's primary function is to produce the rationale.