MooreMuaMu/qwen35-27b-ancient-rl-r32-step250
MooreMuaMu/qwen35-27b-ancient-rl-r32-step250 is a 27 billion parameter Qwen3.5-based language model, fine-tuned using Reinforcement Learning (GRPO) with a LoRA rank of 32. This specific checkpoint (step 250) was selected for its statistically significant improvement over the base model in a held-out evaluation. It is particularly optimized for tasks involving Uyghur, traditional Mongolian, and Tibetan translation and annotation.
Loading preview...
Model Overview
MooreMuaMu/qwen35-27b-ancient-rl-r32-step250 is a 27 billion parameter Qwen3.5-based language model, representing a full merged checkpoint from an experimental Reinforcement Learning (RL) run. It was fine-tuned using the GRPO algorithm with a LoRA rank of 32 and a dropout of 0.05, targeting all linear modules in the Qwen3.5 language model.
Key Characteristics & Performance
This model is the recommended checkpoint from the "Ancient Swift RL Stage 2 R32" run, specifically step250. It demonstrated statistically significant improvements over its base model in a held-out 300-prompt evaluation, with positive deltas across metrics like Exact match, Char-F1, BERTScore-F1, and Semantic composite. Notably, it showed a significant increase in SacreBLEU-2 scores for both corpus and sentence mean deltas.
Training Focus
The training data for this RL fine-tuning primarily consisted of Uyghur, traditional Mongolian, and Tibetan translation and annotation. This specialized dataset and RL approach suggest an optimization for tasks within these specific linguistic and cultural contexts.
Usage Considerations
Users can load this model directly using transformers with Qwen3_5ForConditionalGeneration. It's important to note that later checkpoints from the same run (step300, step350, etc.) showed degraded or collapsed quality, making step250 the optimal choice. As an experimental merged checkpoint, task-specific evaluation is recommended before production deployment.