Overview

The AQuarterMile/WritingBench-Critic-Model-Qwen-7B is a specialized language model derived from the Qwen2.5-7B-Instruct architecture. With 7.6 billion parameters, its primary function is to act as an evaluator for writing tasks, providing critical assessment of generated text.

Key Capabilities

• Writing Evaluation: Fine-tuned on a 50K supervised fine-tuning (SFT) dataset specifically for assessing written responses.
• Criterion-based Scoring: Independently assigns scores on a 10-point scale for each evaluation criterion.
• Justification Generation: Provides detailed justifications alongside each score, explaining the reasoning behind the evaluation.
• Research-backed: Part of the broader WritingBench project, as detailed in the associated paper and GitHub repository.

Training Details

The model was trained using a learning rate of 7e-06, with a total batch size of 64 (1 train_batch_size with 8 gradient_accumulation_steps across 8 devices) over 3 epochs. It utilized the AdamW_TORCH optimizer and a cosine learning rate scheduler with a 0.1 warmup ratio.

Good For

• Automated assessment of written content.
• Providing structured feedback on text generation.
• Research and development in AI-driven writing evaluation systems.