MooreMuaMu/qwen35-27b-ancient-rl-r32-step100

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

MooreMuaMu/qwen35-27b-ancient-rl-r32-step100 is a 27 billion parameter Qwen3.5-based language model, fine-tuned using Reinforcement Learning (RL) with a rank-32 LoRA adapter merged into the base model. This bfloat16 model is specifically optimized for tasks involving ancient languages, showing improvements in SacreBLEU2 scores for Traditional Mongolian, Tibetan, and Uyghur annotation and translation. It is designed for direct loading via the `transformers` library for applications requiring specialized ancient language processing.

Loading preview...

Model Overview

MooreMuaMu/qwen35-27b-ancient-rl-r32-step100 is a 27 billion parameter Qwen3.5-based language model that has undergone Reinforcement Learning (RL) fine-tuning. This model integrates a rank-32 LoRA adapter, which has been merged into the Qwen3.5-27B ancient-language stage-2 base model, and is provided as a full bfloat16 checkpoint. It is ready for direct use with the transformers library, including safetensors shards and necessary tokenizer files.

Key Capabilities & Performance

This model is specifically tailored for tasks involving ancient languages, with evaluations focusing on Tibetan, Traditional Mongolian, and Uyghur. Key performance indicators include:

  • SacreBLEU2 Improvement: The model demonstrates notable improvements in SacreBLEU2 scores, particularly for Traditional Mongolian annotation (+16.77 pp Corpus BLEU-2) and overall Corpus BLEU-2 (+5.56 pp) compared to its base model.
  • Semantic Metrics: While overall semantic composite metrics show a small positive delta (+0.0100), the confidence interval crosses zero, indicating that semantic improvements are not statistically significant across all metrics.
  • Answer Extraction: Maintains a high rate of nonempty extracted answers (around 94-100%) across different ancient language buckets.

Intended Use Cases

This model is suitable for developers and researchers working on:

  • Ancient Language Processing: Specifically for tasks like annotation and translation in Tibetan, Traditional Mongolian, and Uyghur.
  • Research in RL Fine-tuning: Provides a merged checkpoint for studying the effects of RL fine-tuning on specialized language models.
  • Direct Integration: Designed for straightforward integration into projects using the Hugging Face transformers library.

Finetunes
2 models