rishanthrajendhran/POLARIS-no-HRI-9B
The rishanthrajendhran/POLARIS-no-HRI-9B is a 9 billion parameter creative writing model, an ablation variant of POLARIS-9B, developed by Rishanth Rajendhran. It is fine-tuned using the GRPO training recipe with a structured Story Quality reward, specifically optimized for long-form story generation and creative writing tasks. This model excels at generating narrative content, offering substantial improvements over its base Qwen3.5-9B model, particularly for in-distribution lengths up to 4,000 words.
Loading preview...
POLARIS-no-HRI-9B: Creative Writing Model
POLARIS-no-HRI-9B is a 9 billion parameter model developed by Rishanth Rajendhran, serving as an ablation variant of the POLARIS-9B model. It utilizes the same GRPO training recipe, structured Story Quality reward, and hyperparameters as POLARIS-9B, but notably omits human-reference injection (HRI) during training. Instead of incorporating human-written stories, it relies solely on six policy rollouts per group.
Key Capabilities and Performance
This model is a strong creative writing tool, significantly outperforming the base Qwen3.5-9B model. While it performs comparably to POLARIS-9B at in-distribution lengths (1–4k words), its story quality and length adherence show a steeper degradation at longer, out-of-distribution lengths (8–12k words). For instance, its Story Quality score at 8–12k words is 37.7 compared to POLARIS-9B's 44.1, and its length adherence is 0.70 versus 0.72. However, on OOD benchmarks like WritingBench, LongBench-Write, and EQ-Bench Creative, its performance is nearly identical to POLARIS-9B.
Intended Use Cases
- Long-form story generation: Ideal for short stories, flash fiction, and narrative scenes.
- Creative writing: Suitable for essays, book reviews, and podcast scripts.
Limitations
The primary limitation is a steeper quality degradation for very long generations (above 6k words) compared to POLARIS-9B. For use cases requiring extensive long-form content, POLARIS-9B is recommended. It is not intended for factual, knowledge-intensive, legal, medical, or financial content.